Spatial multi omic profiling maps hypoxia-driven pro-fibrotic SPP1⁺ macrophages underpinning immune escape and TACE resistance in hepatocellular carcinoma

Abstract

Background

To elucidate the cellular and molecular mechanisms underlying therapeutic resistance (refractoriness) following transarterial chemoembolization (TACE) in hepatocellular carcinoma (HCC) by comprehensively characterizing the post-treatment tumor microenvironment (TME).

Methods

We employed an integrative spatial multi-omics strategy, combining bulk and single-cell RNA sequencing, subcellular-resolution spatial transcriptomics, and spatial proteomics on tissues from TACE-treated and treatment-naïve HCC patients. Public datasets were used for prognostic and predictive validation, and key findings were confirmed with multiplex immunofluorescence and in vitro experiments.

Results

TACE induced a profoundly hypoxic TME, which drove the selective enrichment of a pro-fibrotic tumor-associated macrophage (TAM) population characterized by high SPP1 expression (SPP1 + TAMs). Spatial mapping demonstrated that these SPP1 + TAMs localize to hypoxic tumor cores, where they remodel the extracellular matrix by producing fibronectin (FN1), establish fibrotic niches, and restrict T-cell activation, thereby orchestrating an immune-excluded microenvironment. The abundance of this macrophage subtype was a robust predictor of TACE resistance.

Conclusions

TACE-induced hypoxia promotes a macrophage-driven fibrotic program that is a key determinant of immune evasion and treatment failure in HCC. Targeting this SPP1 + TAM-mediated fibrotic niche presents a potential therapeutic strategy to overcome TACE resistance and improve clinical outcomes.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s40164-026-00820-1.

Background

Hepatocellular carcinoma (HCC), the most common primary liver malignancy, remains a leading cause of cancer-related mortality worldwide [1]. In HBV-endemic regions like East Asia, it is often diagnosed at advanced stages and high frequency of recurrence with limited curative options [2, 3]. Transarterial chemoembolization (TACE) is commonly employed in clinical practice, particularly for patients with intermediate-stage disease or as a bridging or downstaging strategy for unresectable HCC [4, 5]. However, its therapeutic efficacy is often constrained by high rates of tumor resistance and suboptimal long-term outcomes. Basically, TACE induces ischemic necrosis and hypoxia within the tumor bed, reshaping the TME in ways that can paradoxically promote tumor progression, highlighting the need to elucidate the mechanisms underlying treatment resistance [6–8].

Recent studies have shown that combining TACE with immune checkpoint inhibitors (ICIs) and anti-VEGF or TKI therapies yields promising efficacy and manageable safety in advanced HCC [9, 10]. EMERALD‑1 and LEAP‑012 phase III studies further validate that TACE combined with target therapy and immunotherapy (T + I) prolongs progression-free survival (PFS) relative to TACE alone [11, 12]. These findings suggest that integrating TACE-induced ischemia with systemic immune-vascular modulation can counteract treatment-induced immune evasion, thereby promoting response conversion and widening the therapeutic window for curative interventions. A thorough understanding of TACE-induced TME remodeling is essential for uncovering resistance mechanisms and guiding the rational development of synergistic combination therapies.

Advances in single-cell and spatial multi-omics technologies have begun to illuminate the complex remodeling of the TME following TACE. TACE-induced expansion of TREM2⁺ tumor associated macrophages (TAMs) enhances PD-L1 expression of endothelial and restricts CD8⁺ T-cell infiltration, thereby suppressing antitumor immunity [13]. Cross talk between malignant cells, neutrophils, and CD8⁺ T cells further implicates NABP1⁺ tumor cells as potential drivers of TACE resistance [14]. Despite these progresses, current approaches lack spatial high resolution to fully capture pathological changes post-TACE treatment, underscoring integrative spatial analyses to precisely map treatment-induced alterations in tissue architecture and cellular organization [15–17].

In this study, we applied an integrative spatial multi-omics strategy combining single-cell RNA sequencing (scRNA-seq), spatial transcriptomics and spatial proteomics in both primary tumor and surgically resected specimens under TACE to comprehensively map the topological remodeling after TACE. We identified hypoxia-driven accumulation of SPP1⁺ macrophages highly expressing pro-fibrotic molecules of FN1 as a critical effector mediating extracellular matrix remodeling, fibrosis, and immune evasion. These findings offer a mechanistic basis for TACE-induced immune evasion and tumor progression, highlighting potential strategies for rational combination therapies targeting the TME in HCC.

Results

Single-cell transcriptomic analysis after TACE reveals an immune-exhausted TME

To characterize TACE-induced immune alterations, we integrated multiple datasets, including a publicly available single-cell RNA sequencing dataset related to TACE (TACE-scRNA), [13] in-house spatial transcriptomic (TACE-spatialT) and spatial proteomic datasets (TACE-spatialP). Specifically, the TACE-scRNA dataset consisted of samples from 5 pre-TACE and 5 post-TACE patients (non-paired); the TACE-spatialT dataset included 7 treatment-naïve and 9 post-TACE patients; and the TACE-spatialP dataset comprised 4 treatment-naïve and 5 post-TACE patients. In addition, tumor samples from 5 responders and 5 non-responders after TACE were analyzed by multiplex immunofluorescence (mIF) to validate response-predictive biomarkers (TACE-SpatialM). Tumor core (T), invasive boundary (B), and peritumoral (N) regions were manually annotated by certificated pathologists, while surgically resected formalin-fixed, paraffin-embedded (FFPE) samples were constructed into tissue microarrays (TMAs) for spatial multi-omics profiling. We further incorporated bulk transcriptomic data from TCGA-LIHC (370 samples), ICGC-LIRI (240 samples), HCC-PLANET (250 samples) and a TACE-treated cohort with efficacy assessment (TACE-GEO, 147 samples) (Fig. 1A). Besides, three retrospective clinical cohorts: TACE-neo (N = 60), TACE-resection (N = 75) and TACE-paired (N = 18) were included for validation. This integrative spatial multi-omics analysis enables a refined understanding of TACE-induced topological remodeling in HCC (Supplementary Figure S1A).

Fig. 1.

Single-cell analysis reveals immune remodeling in the HCC TME following TACE treatment. A Schematic overview of the study design integrating public scRNA-seq, bulk RNA-seq, and in-house spatial multi-omics data to investigate TACE-induced immune alterations. B UMAP visualization of major immune cell types identified from TACE-scRNA at the lineage level of primary and post-TACE HCC samples. C Ro/e heatmap of immune cell types across primary (PT) and post-TACE (TT) conditions. D UMAP plot displaying 11 myeloid cell subsets, including macrophages, monocytes, and dendritic cells (DC). E Ro/e heatmap of different myeloid subsets between primary and post-TACE HCC samples. F UMAP plot showing T/NK cell subclusters, including CD8⁺, CD4⁺, and NK cell subsets. G Ro/e heatmap of different T/NK subsets between primary and post-TACE HCC samples. H Heatmap of pathway enrichment scores between different CD8 + T cell subsets. I Violin plots of PDCD1, LAG3 in CD8⁺T_C1_NK-like_GNLY cells between primary and post-TACE groups. J Cytotoxicity and activation score in CD8⁺T_C1_NK-like_GNLY cells between primary and post-TACE groups

Briefly, a total of four major immune components were identified in TACE-scRNA: T/NK cells, myeloid cells, B cells, and plasma cells since TACE-scRNA only contains immune cells (Fig. 1B, Supplementary Figure S1B-D, Supplementary Table 1). Notably, we found myeloid cells exhibited a marked expansion following TACE treatment (TT) compared with pre-treatment (PT), consistent with the findings reported in the original study [13] (Fig. 1C). Further sub clustering analysis of myeloid lineage identified five distinct macrophage populations, two monocyte subsets, and four dendritic cell (DC) subsets (Fig. 1D, Supplementary Figure S1E, Supplementary Table 2), with substantial compositional and transcriptional changes observed post-TACE. Macro_C1_C1QC, Macro_C2_SPP1, and Macro_C4_MT1G subsets were enriched following TACE, with Macro_C2_SPP1 further exhibiting transcriptional features indicative of functional activation (Fig. 1E, Supplementary Figure S1F, G). In contrast, Macro_C5_ISG15, Macro_C6_MKI67, and Mono_C1_CD14 subsets decreased, indicating reduced inflammatory and proliferative myeloid cells after TACE treatment. Classical dendritic cell subsets cDC_C2_CLEC9A and cDC_C3_LAMP3 showed a marked decline, reflecting impaired antigen presentation capacity, while plasmacytoid dendritic cells (pDCs) also decreased, potentially weakening antiviral and immune regulatory functions (Fig. 1E). In parallel, we observed distinct transcriptional profiles and TACE-induced dynamics across T/NK cell subtypes, which were further resolved into five CD8⁺ T cell, four CD4⁺ T cell and two NK cell subsets (Fig. 1F, Supplementary Figure S1H, Supplementary Table 3). We found CD4⁺T_C1_Naive_IL7R and CD4_C3_Treg_FOXP3 were expanded after TACE (Fig. 1G, Supplementary Figure S1I), reflecting a shift toward a less differentiated or early activation state within the CD4⁺ T-cell compartment. IL7R (CD127) expression is typically associated with naïve or memory precursor T cells with proliferative potential, suggesting that TACE-induced tissue injury and antigen release may promote the recruitment or expansion of these populations. In contrast, the relatively modest change in Treg cells may indicate a limited early immunosuppressive response. For CD8 T cells, we found an expansion of CD8_C4_Proliferating_MKI67 and CD8_C5_MAIT_SLC4A10, all of which showed enrichment of hypoxia-related pathways like glycolysis, oxidative phosphorylation, NF-κB signaling, and cellular stress responses (Fig. 1H, Supplementary Fig. 1J). In contrast, two NK cell subsets and CD8_C1_NK-Iike_GNLY involved in cytokine production and immune activation were reduced following TACE, indicating suppression of innate immune functions (Fig. 1H). Compared with pre-TACE tumors, the main effector CD8⁺ T cell subset, CD8_C1_NK-like_GNLY cells showed increased expression of exhaustion markers such as PDCD1 and LAG3, along with reduced cytotoxicity and activation, indicating a transition toward a dysfunctional state after treatment (Fig. 1I, J, Supplementary Fig. 1K). Taken together, we found transcriptomic signatures suggesting immune suppression in the TME, indicating the underlying TACE refractory mechanism of HCC.

Integrated transcriptomic analyses identify SPP1⁺ TAMs associated with poor prognosis and TACE refractoriness in HCC

Totally, we used TACE-scRNA data during the treatment of TACE to characterize the whole immune landscape of HCC at a more defined resolution compared with Fig. 1B (Fig. 2A, Supplementary Figure S2A). To further identify key immune cell subpopulations associated with clinical outcomes in HCC, we integrated TACE-scRNA dataset with bulk transcriptomic data from TCGA-LIHC using the scPAS algorithm [18]. We systematically mapped prognosis-associated cell types and identified three subpopulations most strongly correlated with patients’ survival: CD8_C4_Proliferating_MKI67, Macro_C6_MKI67 and Macro_C2_SPP1 (Fig. 2B, Supplementary Figure S2B, C). These results indicate that proliferative T cells and specific macrophage subsets may play key roles in determining HCC prognosis. To further validate these associations, we performed immune cell deconvolution on ICGC-LIRI samples using the CIBERSORT algorithm (Supplementary Figure S2D) [19]. We found that higher proportions of Macro-C1-C1QC, Macro-C3-PDGFC, and Macro-C4-MT1G were significantly associated with better overall survival (OS), whereas Macro_C2_SPP1 and Macro-C6-MKI67 were linked to worse OS (Fig. 2C, Supplementary Figure S2E). Pseudo-time trajectory analysis of macrophages in TACE-scRNA revealed that Macro_C2_SPP1 and Macro-C6-MKI67 were positioned in an intermediate state along the differentiation path (Fig. 2D, Supplementary Figure S2F, G). To investigate treatment-specific immune dynamics, we analyzed a publicly available dataset of HCC patients with annotated responses after TACE from GEO database (TACE-GEO). Differential gene expression analysis between responders (R) and non-responders (NRs) revealed distinct transcriptional programs (Fig. 2F, G, Supplementary Figure S2H, Supplementary Table 4). Responders were enriched for pathways related to fatty acid metabolism and immune activation, whereas non-responders exhibited upregulation of hypoxia, glycolysis, and collagen formation pathways (Fig. 2H). These results highlight divergent metabolic and stromal remodeling responses that may underlie TACE efficacy. Deconvolution of macrophage subtypes using TACE-scRNA as a reference in TACE-GEO further showed Macro-C1-C1QC, Macro-C3-PDGFC, and Macro-C4-MT1G were significantly enriched in responders, while Macro_C2_SPP1 and Macro-C6-MKI67 were predominant in non-responders (Fig. 2I). This consistency with ICGC-LIRI cohort underscores the robustness of these subtypes as predictive biomarkers for both survival and treatment response. Interestingly, we also observed a significant enrichment of the cDC-C2-CLEC9A dendritic cell subset in responders (Fig. 2I), which was markedly reduced following TACE treatment (Fig. 1E). This suggests that while cDC-C2-CLEC9A cells may contribute to treatment response via antigen presentation or immune priming, TACE may simultaneously deplete this beneficial population potentially to limit long-term immune activation.

Fig. 2.

Identification and validation of prognosis of SPP1⁺Macrophages in HCC after TACE treatment. A Overall UMAP visualization of immune cell subtypes from integrated TACE-scRNA dataset. B Boxplot of Prognosis-associated risk score identified using the scPAS algorithm in the TCGA-LIHC cohort. C Kaplan–Meier survival curves of percent of Macro_C2_SPP1 with patient prognosis. D Pseudotime trajectory of different subtypes of macrophage differentiation. E Pseudotime of different subtypes of macrophages. F Principal component analysis of data with annotated TACE response status used to assess treatment-specific immune dynamics. G Volcano plot of differentially expressed genes (DEGs) between responders and non-responders. H GSEA enrichment of DEGs between responders and non-responders. I Violin plot of different cell types deconvoluted between responders and non-responders. J Feature importance plot from XGBoost model identifying top predictors of TACE non-response. K Deconvolution analysis of PLANET dataset. L Violin plot summarizing predictive cell subtypes associated with TACE recurrence risk in PLANET dataset

Finally, we trained an XGBoost machine learning model using subcellular deconvolution features to predict TACE response. Among all subpopulations, we found Macro_C2_SPP1 and Macro-C6-MKI67 emerged as the top two predictors of non-responders (Fig. 2J, Supplementary Figure S2I). These results further support the potential clinical utility of these immunosuppressive macrophage subsets to enhance the therapeutic benefit of TACE. Deconvolution analysis of the PLANET dataset [20] also revealed that CD8-C5-MAIT-SLC4A10, Macro-C4-MT1G, cDC-C2-CLEC9A, and cDC-C1-CD1C were enriched in patients of no relapse after TACE, whereas Macro-C2-SPP1 was associated with recurrence, suggesting their potential as predictive markers for TACE response and recurrence risk (Fig. 2K, L, Supplementary Fig. 2 J). Collectively, these multi-cohort analyses consistently identified Macro-C2-SPP1 as a key myeloid subtype associated TACE resistance, highlighting it as a potential target for further functional investigation.

Single-cell spatial transcriptomic mapping of HCC reveals high-resolution tissue organization after TACE treatment

To investigate the impact of TACE on topological reorganization in HCC, we collected surgically resected FFPE tumor specimens from 9 patients who underwent downstaging treatment of TACE and compared them with primary tumor samples from 7 patients (Supplementary Table 5). Comparative analysis showed no significant differences between the treatment-naïve group and the post-TACE group in several key clinical parameters, including tumor stage, presence of cirrhosis, sex, age, and microvascular invasion (MVI). Post-treatment tumor samples were obtained at a relative variable timepoints following the last TACE session, including both early-phase samples (within 6 weeks) and later-phase samples (beyond 6 weeks). Spatial single-cell transcriptomic profiling was conducted using the CosMx Spatial Molecular Imaging (SMI) platform with a 1000-gene panel for high-resolution in situ cellular characterization (Supplementary Table 6), while a subset of samples (5 TACE-treated patients and 4 treatment naive patients) simultaneously underwent spatial proteomic analysis to yield complementary protein-level information (Supplementary Table 7). Integration of single cell spatial transcriptomic and proteomic data enabled the construction of a comprehensive atlas capturing TACE-induced spatial remodeling of the HCC microenvironment.

Following machine learning assisted cell segmentation and quality control, we identified a total of 217,825 single cells with spatially resolved transcriptomic profiles (Supplementary Figure S3A, B). Functional annotation categorized all cells into seven major lineages, including hepatocytes, malignant cells, cholangiocytes, T/NK cells, B/plasma cells, fibroblasts, myeloid cells and endothelial cells (Supplementary Figure S3C, D). Further subclustering identified six lymphocyte subsets, including CD8⁺ effector T cells (CD8⁺ Teff), CD8⁺ exhausted T cells (CD8⁺ Tex), CD4⁺ T cells, regulatory T cells (Tregs), B cells, and plasma cells; three myeloid populations comprising monocytes/macrophages (Mono/Mph), neutrophils (Nph), and dendritic cells (DCs); five subtypes of cancer-associated fibroblasts (CAFs), including antigen-presenting CAFs (aCAFs), myofibroblastic CAFs (mCAFs), vascular CAFs (vCAFs), inflammatory CAFs (iCAFs) and pericytes; and two endothelial subtypes including capillary and lymphatics (Fig. 3A-C, Supplementary Table 8), covering primary tumor (PT) and TACE-treated tumor (TT) sample (Supplementary Figure S3E). Cell-type correlation analysis confirmed the robustness and accuracy of cell identity assignment (Fig. 3D). We found different samples exhibited a variation of cell types, whereas malignant and hepatocytes were the most frequent cell types in the TME, reflecting a more accurate cell type distribution in the TME compared with scRNA-seq data (Fig. 3E, F, Supplementary Figure S3F). Interpretation of an immunosuppressive TME after TACE is not only based on a combined assessment of cell-type composition but also their spatial distribution. By mapping cell coordinates back to their original spatial locations, we reconstructed a subcellular resolution spatial atlas of HCC tissue post-TACE (Fig. 3G). We observed an enrichment of CD8 + Teff cells, B/Plasma cells, DCs and different CAFs at the invasive boundary, forming a spatial pattern suggestive of a physical and stromal barrier. Besides, capillaries and pericytes were predominantly localized within the tumor core, while neutrophils and lymphatics were more abundant in peritumoral regions (Fig. 3H). TACE treatment induced an immunosuppressive and pro-angiogenic TME, characterized by increased abundance of CD8⁺ Tex cells, Tregs, capillaries, pericytes, and vCAFs, indicating TACE may promote neovascularization and immune evasion to contribute to tumor progression (Fig. 3I).

Fig. 3.

Single-cell spatial transcriptomic and proteomic profiling reveal TACE-induced spatial remodeling in HCC. A Functional annotation of identified single cells from CosMx SMI spatial transcriptomics data, categorized into different cell types. B Heatmap of marker genes of different cell types. C UMAP plot of marker genes in different lineages. D Cell-type correlation matrix confirming the robustness and accuracy of cell identity assignment across samples. E Variation in cell type composition across samples. F Stack barplot of different cell types across Tumor (T), invasive boundary (B) and adjacent normal (N) within both primary tumor (PT) and post-TACE tumor (TT). G Spatial mapping of single-cell coordinates back onto tissue sections, reconstructing a high-resolution spatial atlas of HCC tissue architecture post-TACE treatment. H Ro/e heatmap of different cell types across T, B and N within both PT and TT. I Boxplot of relative frequency of different cell types between PT_T and TT_T. J H&E and matched IF images illustrating the Regions of Interests (ROI) selection in GeoMx spatial proteomics data. K Quantitative digital pathology analysis showing higher proportions of immune cells in peritumoral regions and enrichment of fibrotic stroma within tumor regions. L Heatmap of differentially expressed proteins between TACE-treated and treatment-naïve tumor regions. M Enrichment analysis of differentially expressed proteins demonstrating significant activation of hypoxia-related signaling pathways in tumor regions following TACE treatment

To further analyze the dynamic changes induced by the TACE treatment at the protein level, we performed GeoMx spatial proteomic profiling on 5 TACE-treated patients (a total of 31 ROIs) and 4 treatment naive patients (a total of 18 ROIs) from the spatial transcriptomic cohort (Fig. 3J). Using DAPI, CD45, CK8/18, and α-SMA as morphological markers to select ROIs, we firstly quantified the relative proportion of immune cells, epithelial cells, and fibrotic stroma compartments. Peritumoral areas showed a higher proportion of immune cells, whereas fibrotic stroma was predominantly enriched within tumor regions (Fig. 3K). Focusing on tumor regions, we identified multiple proteins upregulated after TACE treatment (Fig. 3L), while differential and enrichment analyses revealed a significant upregulation of hypoxia-related signaling in tumors post-treatment. (Fig. 3M, Supplementary Table 9). In summary, our integrated spatial multi-omics analyses provide a molecular insight of TACE-induced TME changes that may underlie therapeutic resistance.

SPP1⁺ TAMs induced by hypoxia after TACE are enriched in the tumor core region

Spatial multi-omics mapping facilitates the elucidation of the critical role of SPP1⁺ macrophages in mediating resistance to TACE treatment as previously described. Firstly, macrophages from spatial single-cell transcriptomic data were stratified into SPP1⁺ and SPP1⁻ populations based on cluster-level SPP1 expression (Fig. 4A). Specifically, we examined SPP1 expression across all Mono/Mph subclusters and examined distribution of SPP1 to defined clusters with relatively high SPP1 expression as SPP1⁺ macrophages, whereas clusters with low or negligible expression were classified as SPP1⁻ macrophages (Supplementary Fig. 3G, H). Signature scoring further confirmed that C1 and C4 clusters showed the highest SPP1-associated transcriptional activity (Supplementary Fig. 3I). We found SPP1⁺ macrophages were preferentially localized to the tumor core, and their abundance markedly increased following TACE (Fig. 4B). Spatial profiling revealed a pronounced gradient of hypoxia signaling, increasing from the peritumoral region to the tumor core, which intensified after TACE (Fig. 4C). Spatial in situ validation further confirmed the co-localization of SPP1⁺ macrophages and high hypoxia activity in tumor regions post-treatment, indicating that SPP1⁺ TAM enrichment is spatially aligned with hypoxic niches (Fig. 4D, E). To identify robust molecular features of SPP1⁺ macrophages, we intersected differential expressed genes (DEGs) from TACE-spatialT and TACE-scRNA and found a consistent upregulation of pro-fibrotic and hypoxia-induced glycolysis genes that existed in TACE-spatialP panel, including SPP1, FN1, LGALS3, ITGB1 and LDHA, highlighting their role in matrix remodeling and fibrosis (Fig. 4F, Supplementary Table 10). Spatial proteomics data also confirmed that these proteins were elevated after TACE treatment in tumor cores (Fig. 4G). Signaling activity scoring across both TACE-spatialT and TACE-scRNA data demonstrated strong enrichment of hypoxia phenotypes within SPP1⁺ macrophages [21] (Fig. 4H-J) and inferred HIF1A as a critical transcription factor (TF) for this subtype [22] (Fig. 4, K-M). Together, these findings provide converging spatial and molecular evidence that the accumulation and functional activation of SPP1⁺ macrophages following TACE is featured by hypoxia.

Fig. 4.

TACE exacerbates hypoxia in the tumor core region and recruits SPP1⁺macrophages. A Identification of SPP1⁺ macrophages within the CosMx monocyte/macrophage (Mono/Mph) subpopulation based on SPP1 expression. B Ro/e analysis showing the spatial distribution patterns of SPP1⁺ Macrophages versus SPP1⁻ Macrophages. C Boxplot of hypoxia activity in SPP1⁺ macrophages across T, B and N within both PT and TT. D Spatial distribution of SPP1⁺ and SPP1⁻ Macrophages across T, B and N within both PT and TT. E Spatial mapping of hypoxia scores across T, B and N within both PT and TT. F Venn diagram illustrating the intersection of SPP1⁺ Macrophage marker genes identified from CosMx and scRNA-seq data with the GeoMx protein data. G GeoMx proteomic data showing the expression of intersecting genes in tumors across T and N within both PT and TT. H Heatmap of PROGENy pathway activity for various cell types in the CosMx dataset. I Heatmap of PROGENy pathway activity for each myeloid subpopulation in the scRNA-seq dataset. J HALLMARK HYPOXIA UCell score for each myeloid subpopulation in the scRNA-seq dataset. K CollectTRI analysis showing the activity of transcription factors for each myeloid subpopulation in the scRNA-seq dataset. L Boxplots demonstrating increased HIF1A expression and a higher HYPOXIA score in the Macro_C2_SPP1 subset following TACE treatment. (M) GO functional enrichment analysis of marker genes for the Macro_C2_SPP1 subset

SPP1⁺ TAMs shape a spatial pro-fibrotic niche to facilitate immune escape

To characterize the spatial heterogeneity of tumor cells within TME, we extracted all cells and performed unsupervised clustering, identifying seven malignant subpopulations and one cluster of normal hepatocytes (Fig. 5A, B, Supplementary Table 11). Then, we applied consensus non-negative matrix factorization (cNMF) to tumor cells derived from TACE-SpatialT data to identify eight recurrent transcriptional programs (Supplementary Figure S4A, B). This analysis resolved a set of gene modules representing distinct functional states of malignant cells (Supplementary Table 12). Program 1 was enriched for dendritic cell chemotaxis related genes, whereas Program 2 was linked to glycolysis and glucose metabolic processes. Program 3 was characterized by signatures of oxidative stress, epithelial and fibroblast proliferation, while Program 4 was associated with cell chemotaxis and lymphocyte proliferation. Program 5 reflected apoptotic pathways, and Program 6 was enriched for genes involved in monocyte differentiation. Program 7 highlighted extracellular matrix remodeling, and Program 8 was primarily related to kinase activity. Together, these programs delineate diverse functional states of tumor cells, spanning immune interaction, metabolism, stress adaptation, and microenvironmental remodeling (Fig. 5C, D). Spatial niche analysis using the Spatopic algorithm [23] revealed 15 distinct spatial ecosystems within the TME (Fig. 5E, F). We found Niche 12 was dominated by B/plasma cells and CAFs, while niches 11, 13, and 14 were enriched in neutrophils and lymphatic endothelial cells. Niche 1 was primarily composed of mCAFs and cholangiocytes, and niche 7 represented normal hepatocytes. The remaining niches were predominantly composed of various tumor cell subpopulations. Conversely, niche 12 showed a marked reduction in T, B, and N regions post-treatment, suggesting depletion of stromal and B cell–enriched environments (Fig. 5G). Notably, niche 2 was composed of dendritic cells, CD8⁺ effector and exhausted T cells, CD4⁺ T cells, SPP1⁺ macrophages, and CXCL10⁺ tumor cells. Following TACE treatment, niche 2 was significantly enriched in both tumor (T) and border (B) regions (Fig. 5H). CellChat analysis revealed that interactions between SPP1⁺ macrophages and other T cell subtypes were strengthened via FN1-CD44 axis (Fig. 5I), reflecting the immune suppression mediated by this pro-fibrotic macrophage [24–26].

Fig. 5.

Spatially resolved transcriptomics deciphers tumor heterogeneity and immunosuppressive niches mediated by SPP1⁺ macrophages post TACE. A Identifying various tumor subpopulations and normal hepatocytes from CosMx spatial data. B Expression of marker genes for tumor subpopulations and hepatocytes. C 8 distinct gene programs of CosMx data in tumor cells using cNMF analysis. D Feature plots of activity score for each cNMF gene program. E Spatial distribution of 15 niches identified by Spatopic algorithm. F Heatmap of relative enrichment of different cell types in 15 different niches. G Stacked bar plot showing the cellular composition of each identified niche. H Heatmap displaying the Ro/e for the enrichment of each niche across T, B and N within both PT and TT. I CellChat analysis identifying FN1-CD44 interaction between SPP1⁺ Macrophages and T cell subtypes

SPP1⁺ macrophages drive immune evasion and TACE resistance mediated by FN1

To further investigate the dynamic changes of SPP1⁺ macrophages during the treatment of T + I, we reanalyzed a published single-cell RNA sequencing dataset from HCC patients [27]. Subclustering of immune cells identified the Macro_C03_TREM2 population were highly enriched SPP1 and FN1 (Fig. 6A, B), mainly involved in regulation of T cell activation and hypoxia (Fig. 6C). We found this subset showed a marked reduction after T + I therapy (Fig. 6D) and was further decreased in responders compared to non-responders (Fig. 6E). Hypoxia-conditioned macrophages induced SPP1 expression and collagen remodeling in LX-2 cells, supporting their pro-fibrotic role (Supplementary Fig. 4C). FN1 is often regarded as an extracellular matrix protein and is mainly expressed by fibroblast [28, 29]. In our study, we found an enrichment and colocalization of FN1 and CD68 in both CosMx spatial data and multiplexed immunohistochemistry (mIHC) slides (Supplementary Fig. 4D), supporting the interaction between fibrotic and macrophage compartments in the post-TACE TME. ³⁰ Meanwhile, LGALS3, a β-galactoside-binding lectin in the galectin family has been identified as a novel immunosuppressive target in pancreatic cancer and exert significant effects on multiple key immune cell subsets, including T cells, B cells, and antigen-presenting cells (APCs) [31, 32]. Notably, TACE-refractory and pro-fibrotic signatures including SPP1, FN1, and LGALS3 were significantly enriched in non-responders following T + I (Fig. 6F), while FN1 knockdown using siRNA targeting FN1 in PMA-induced THP-1 cells (Supplementary Fig. 4E-G) led to a pronounced reduction of LGALS3 expression, suggesting that FN1 may positively regulate this lectin (Supplementary Fig. 4H, I). In line with this, transcriptomic profiling revealed that loss of FN1 was accompanied by robust activation of T cell–associated pathways, implying that the FN1–LGALS3 axis may contribute to immune suppression in HCC (Fig. 6G, H). Hypoxia-conditioned macrophages impaired CD8⁺ T cell cytotoxic function, which was partially rescued by FN1 knockdown (Supplementary Figure S4J).

Fig. 6.

FN1⁺/SPP1⁺macrophages may mediate immune evasion and TACE resistance. A Subclustering of myeloid cells under the treatment of target combined with immunotherapy. B Vlnplot of SPP1 and FN1 expression in different myeloid subtypes. C Functional enrichment analysis of Macro_C03_TREM2. D Heatmap of Ro/e of different myeloid subtypes between pre-treatment (pre_T) and post-treatment (post_T) group. E Heatmap of Ro/e of different myeloid subtypes between post-treatment non-responders (post_T_NR) and post-treatment responders (post_T_R) group. F Vlnplot of SPP1, FN1, LGALS3 between post_T_NR and post_T_R. G Heatmap of DEGs of RNA-seq data of FN1 KD PMA induced macrophages derived from THP-1 cells. H GO enrichment analysis of DEGs of RNA-seq data of FN1 KD PMA induced macrophages derived from THP-1 cells. I UMAP of 14-plex mIF on HCC tumors after TACE treatment identified 12 major immune and stromal populations across R and NR. J Representative mIF images validating identified cell subtypes, the DNA channel was SYTOX Orange. K Heatmap of marker protein enrichment in distinct cell types in post-TACE samples. L Representative mIF images showing increased leukocyte infiltration and extensive necrosis in R compared with NR, the DNA channel was SYTOX Orange. M Quantification of CD163⁺ M2 macrophages in R and NR. N Quantification of FN1⁺SPP1⁺PDCD1⁺ M2 macrophages in R and NR

To further linking the spatial ecosystems and ligand-receptor interactions identified above to clinical outcomes, we next performed a 14-plex spatial multiplex immunofluorescence (mIF) on tumor tissues from 5 TACE responders and 5 non-responders (Supplementary Table 13). Treatment response was assessed according to the modified Response Evaluation Criteria in Solid Tumors (mRECIST), whereby complete response (CR) was defined as disappearance of any arterial enhancement in all target lesions, partial response (PR) as at least a 30% decrease in the sum of diameters of viable target lesions, progressive disease (PD) as at least a 20% increase in the sum of diameters of viable lesions or the appearance of new lesions, and stable disease (SD) as neither PR nor PD. Persistent PD after three or more consecutive standardized TACE sessions was defined as TACE non-response according to the Chinese College of Interventionalists (CCI) consensus criteria. After image-based single-cell segmentation, a total of 298,189 spatially resolved cells were identified and classified into 12 major populations, including CD34⁺ endothelial cells, CD8⁺ T cells, CD68⁺ M1 macrophages, CD163⁺ M2 macrophages, FOXP3⁺ Tregs, FN1⁺ fibroblasts, and TLS-associated subgroups such as CD3⁺CD8⁺CD20⁺, CD4⁺CD20⁺, and CD4⁺CD8⁺ cells. Additional clusters of SPP1⁺ and unknown cells were detected, likely representing tumor or hepatocyte populations given the lack of tumor-specific markers in our panel (Fig. 6I). The existence of these subtypes was further validated by the mIF images (Fig. 6J). We mapped the spatial organization of multiple cell types and signaling components within the TME and correlated these spatial features with clinical response to TACE. Our results show that the identified spatial niche, as well as the interaction axis such as FN1-CD44 are significantly enriched in the TACE non-response group, supporting the notion that these microenvironmental structures may be associated with treatment resistance rather than representing a general post-treatment remodeling phenomenon. Importantly, we identified a distinct FN1⁺SPP1⁺PDCD1⁺ M2 subset, confirming the presence of this macrophage population in post-TACE samples (Fig. 6K). We found a significant infiltration of leukocytes and extensive necrosis in responders after TACE compared with non-responders (Fig. 6L). Quantitative analysis demonstrated that both CD163⁺ M2 macrophages and FN1⁺SPP1⁺PDCD1⁺ M2 macrophages were significantly enriched in non-responders compared with responders (Fig. 6M, N). These results highlight the spatial enrichment of immunosuppressive FN1⁺SPP1⁺PDCD1⁺ M2 macrophages as a key feature distinguishing non-responders from responders to TACE. Besides, we included a TACE-neo cohort which has 60 patients with early-stage (BCLC A/CNLC 1b) resectable HCC who underwent neoadjuvant TACE followed by curative resection, along with a TACE-resection cohort, in which 75 patients with intermediate- to advanced-stage HCC (BCLC stage B/C) who underwent surgical resection following TACE down-staging treatment were included retrospectively. We found SPP1⁺ TAMs were associated with poor TACE response in these two cohorts (Supplementary Tables 15–16, Supplementary Figure S4K-N). Moreover, considering the samples used in our study are not patient-matched (pre and post TACE), we retrospectively reviewed our institutional cohort thoroughly and identified a TACE-paired cohort consisting of 18 patients with matched pre-TACE biopsy specimens and post-TACE surgical resection samples (Supplementary Table 17), due to pre-TACE tumor biopsy is not routinely performed, as treatment decisions are primarily based on radiological diagnosis. We specifically evaluate the association between SPP1⁺ TAM changes within the same patients and TACE response, thereby substantially reducing the potential confounding introduced by inter-patient heterogeneity. We found these TACE-related changes were significantly associated with TACE response (Supplementary Figure S4O-P).

Discussion

TACE remains a cornerstone for intermediate-to-advanced HCC, yet its clinical efficacy is limited by heterogeneous responses and induction of an immunosuppressive TME. Despite causing ischemic tumor necrosis, TACE yields suboptimal long-term outcomes, with reported 1-, 3-, and 5-year overall survival rates of approximately 82%, 47%, and 26% respectively, highlighting the inherent resistance and limitations of TACE monotherapy [33]. Persistent hypoxia, pro-angiogenic rebound, and incomplete embolization may collectively drive tumor persistence and progression [34, 35]. Identifying robust biomarkers of TACE resistance holds promise for patient stratification and for devising rational combination regimens. In this study, by integrating scRNA-seq, bulk RNA-seq, spatial transcriptomics and spatial proteomic data, we demonstrated that localized hypoxia induced by TACE promotes the emergence of a fibrotic and immunosuppressive niche orchestrated by SPP1⁺ TAMs by constructing a fibronectin-rich ECM, accompanied by a paucity of CD8⁺ T cells and restrict effector lymphocyte activation [36, 37].

FN1 (Fibronectin) is a major scaffold of the desmoplastic ECM and serves as a reservoir for chemokines and growth factors within the TME. In fibrotic tumors, FN1-rich extracellular matrices can physically restrict T cell infiltration and immobilize immune-regulatory molecules, thereby contributing to local immunosuppression within TME [38, 39]. LGALS3 (Galectin-3), a carbohydrate-binding lectin capable of cross-linking cell-surface glycoproteins, not only promotes tumor progression but also dampens T cell–mediated immune responses [40, 41]. Importantly, macrophage-derived Galectin-3 can reinforce fibrotic signaling by stabilizing TGF-β receptors on fibroblasts, amplifying TGF-β1-driven ECM deposition [42]. In our study, we observed that FN1 and LGALS3 were significantly enriched in SPP1⁺ macrophages after TACE treatment, and FN1 knockdown in macrophages reduced LGALS3 expression, supporting a feed-forward FN1→Galectin-3 loop. These macrophage-derived ECM components likely exacerbate tissue fibrosis, hinder immune cell trafficking, and limit therapeutic delivery, collectively creating a hypoxic, immune-excluded niche that underlies TACE resistance. By linking macrophage-driven ECM remodeling to both immunosuppression and treatment refractoriness, our findings provide a mechanistic rationale for targeting the FN1–Galectin-3 axis to improve the efficacy of TACE combined with immunotherapy in HCC.

TACE treatment is widely recognized to induce a hypoxic TME by blocking the tumor’s blood supply. Previous studies have primarily highlighted the direct effects of hypoxia on fibroblasts, showing that low oxygen tension can activate fibroblasts and drive tissue fibrosis [43–45]. However, accumulating evidence now indicates that hypoxia also induces functional reprogramming of macrophages, which in turn promotes fibrotic remodeling [46–48]. Single-cell and functional studies have defined an emergent SPP1⁺ macrophage subset that is enriched in hypoxic niches and potently drives fibroblast → myofibroblast conversion via paracrine cues like SPP1, CXCL4 and matrix-modifying programs, positioning SPP1⁺ macrophages as a conserved pro-fibrotic effector across organs [30]. In infection-induced epididymo-orchitis, S100A4⁺ macrophages derived from monocytes acquire myofibroblast-like features and secrete collagen I and fibronectin through TGF-β/STAT3 signaling, directly contributing to fibrotic remodeling [49]. Similarly, in hypertensive cardiac injury, infiltrating macrophages undergo macrophage-to-myofibroblast transition (MMT) via the ALKBH5/IL-11/IL11RA1 axis, stabilizing IL-11 mRNA and promoting fibroblast activation [50]. Notably, hypoxia can also induce antifibrotic macrophage programs, as HIF-1α-dependent oncostatin-M secretion suppresses TGF-β1-driven fibroblast activation during cardiac remodeling [51]. In acute myocardial infarction, hypoxia induces VSIG4 expression in M2 macrophages, which enhances TGF-β1 and IL-10 production and drives cardiac fibroblast-to-myofibroblast conversion, thereby promoting scar formation [51]. Collectively, these findings highlight macrophage plasticity as a central mechanism through which hypoxia drives extracellular matrix remodeling and tissue fibrosis, which in turn facilitates immune evasion.

More broadly, our findings provide a mechanistic rationale for integrating locoregional therapies such as TACE with systemic targeted or immune-based strategies to overcome hypoxia-driven resistance. We demonstrate that TACE-induced hypoxia fosters a fibrotic and immunosuppressive TME through the activation of SPP1⁺ macrophages and ECM remodeling, which not only promotes immune evasion but also limits drug penetration and T cell infiltration. This mechanistic link between hypoxia, macrophage-driven fibrosis, and immune exclusion highlights fibrosis as a key modifiable barrier in TACE-refractory HCC. Combining TACE with anti-fibrotic or vascular-normalizing targeted agents may mitigate hypoxia-induced ECM remodeling, while immunotherapy can restore antitumor immunity by reinvigorating exhausted T cells [5, 52, 53]. Importantly, the identification of macrophage-derived markers such as SPP1 and FN1 offers an opportunity to stratify patients most likely to benefit from such combination approaches. Future biomarker-driven trials validating these candidates are warranted. Conceptually, disrupting the macrophage–FN1–Galectin-3 circuit may reprogram the fibrotic niche into an immune-permissive microenvironment, providing a strong biological foundation for TACE-ICB combinations and potentially improving clinical outcomes in patients with advanced HCC [54–56].

Recent studies have consistently identified SPP1⁺TAMs as an immunosuppressive macrophage population associated with poor prognosis and therapeutic resistance in HCC, including resistance to immune checkpoint blockade [57–59]. Rather than redefining the prognostic significance of SPP1⁺TAMs, our study provides new insights into their functional role in the context of TACE. We identified a profibrotic subset of SPP1⁺TAMs characterized by elevated FN1 and LGALS3, which was preferentially enriched in TACE refractoriness and spatially associated with ECM remodeling and immune exclusion. These findings suggest that TACE resistance is associated not simply with the presence of SPP1⁺TAMs, but with their functional polarization toward a profibrotic remodeling program. We propose that TACE-induced tissue injury promotes the accumulation of these profibrotic macrophages, which subsequently reinforce fibrotic niches and suppress antitumor immunity, thereby contributing to therapeutic resistance. Together, our findings extend previous studies by linking the profibrotic functional state of SPP1⁺TAMs to TACE-specific microenvironmental remodeling and provide a potential mechanistic basis for combining TACE with therapies targeting macrophage-mediated fibrosis or immune suppression.

This study highlights the complementary strengths of spatial transcriptomics, spatial proteomics (GeoMx DSP570), and immunofluorescence-based validation in resolving the TME. Spatial transcriptomics enables unbiased, high-resolution mapping of gene expression patterns across tissue architecture, providing a comprehensive view of cellular states and interactions at single-cell resolution. In contrast, GeoMx DSP570 offers targeted yet highly quantitative protein-level measurements within defined spatial regions, bridging the gap between transcriptional programs and functional protein expression. Immunofluorescence further provides high-resolution visualization and validation at the cellular and subcellular levels, allowing precise localization and co-expression assessment of key markers. Together, these approaches operate across complementary molecular layers and spatial scales, enabling robust cross-validation and a more integrated understanding of the spatial and functional heterogeneity underlying tumor–microenvironment interactions.

Our study has several limitations, including a relatively small cohort of paired TACE-treated specimens, which restrict temporal analyses and the validation of predictive biomarkers. Addressing these gaps will require standardized spatial multi-omics workflows and the integration of more pre- and post-intervention biopsies in prospective trials, enabling robust evaluation of hypoxia- and macrophage-driven fibrotic signatures. Besides, the temporal dimension of post-TACE tissue sampling warrants careful consideration. Emerging evidence suggests that the immune microenvironment undergoes dynamic and stage-specific remodeling following TACE [60]. In this context, the lack of precise annotation of sampling timepoints may introduce heterogeneity in the interpretation of immune composition and transcriptional states. Therefore, future studies should incorporate more rigorous and standardized study designs with clearly defined sampling intervals, ideally through prospective cohort settings. Such approaches will enable more accurate dissection of the temporal evolution of treatment-induced immune responses and improve the mechanistic understanding of TACE-associated immune remodeling and resistance. Moreover, the mechanistic link between SPP1⁺ macrophages and TACE resistance is supported primarily by correlative clinical observations and in vitro functional assays. Although these findings suggest a role for SPP1⁺ macrophages in shaping an immunosuppressive and pro-fibrotic microenvironment, future in vivo studies are required to establish their direct causal contribution to therapeutic resistance.

Another limitation of this study is the inherent procedural heterogeneity of TACE clinically, as efficacy can be influenced not only by tumor-intrinsic biological characteristics but also by operator experience, the degree of superselective catheterization, embolization strategy, and tumor vascular anatomy. Such variability may confound comparisons between TACE responders and non-responders, particularly in relatively small discovery cohorts. To mitigate this limitation, we further validated our findings in two independent cohorts with more clinically homogeneous treatment settings, including patients receiving neoadjuvant TACE before curative resection and a cohort undergoing surgical resection following TACE. The consistent associations observed across these validation cohorts support the robustness and generalizability of our findings, suggesting that the identified spatial immune features are not merely attributable to procedural variability but represent biologically relevant determinants of TACE response.

Preclinical modeling of TACE in HCC remains challenging at present. While relevant HCC mouse models with TACE-like ischemic injury can provide valuable mechanistic insights into treatment-induced hypoxia, immune remodeling, and therapeutic resistance, faithfully recapitulating the clinical procedure in mice is technically difficult due to their small vascular anatomy and the complexity of selective arterial embolization [61]. Larger animal models, such as rats or rabbits, allow more feasible catheter-based embolization and better simulation of clinical TACE; however, these systems are associated with substantially higher costs and lower experimental throughput [62, 63]. These limitations highlight the current trade-off between physiological relevance and experimental feasibility, underscoring the need for continued development of optimized and scalable preclinical models for TACE research.

Importantly, our findings highlight the potential of TACE not only as a locoregional therapy but also as a precision platform for the targeted delivery of immunomodulatory or anti-hypoxia agents. By concurrently mitigating ECM-mediated immune exclusion and reinvigorating antitumor immunity, such combination strategies could overcome TACE resistance and enhance therapeutic efficacy. Future studies validating macrophage-derived biomarkers, such as SPP1, FN1, and Galectin-3, will be critical to guide patient stratification and optimize TACE-based combination regimens in HCC.

Methods

Human tissue samples

FFPE tissue specimens were retrospectively collected from four patients with HCC who received TACE treatment followed by surgical resection at the Department of Pathology, Beijing Tsinghua Changgung Hospital. The study protocol was reviewed and approved by the Ethics Committee of Beijing Tsinghua Changgung Hospital, Tsinghua University (Approval No. 25521-0-02), and written informed consent was obtained from all participants. Tumor cores, invasive margins, and adjacent non-tumorous liver tissues were identified and isolated by board-certified pathologists.

TACE response definition

The definition of treatment response in this study was primarily used to characterize the level of response to TACE and to identify TACE refractoriness. In general, TACE refractoriness refers to a state in which tumors fail to achieve or maintain a therapeutic response after repeated TACE procedures, indicating insensitivity to further TACE treatment. In contrast, recurrence describes tumor regrowth or the emergence of new lesions following an initial response or disease control. Treatment response was assessed according to the consensus criteria of the Chinese College of Interventionalists (CCI) and evaluated based on the modified Response Evaluation Criteria in Solid Tumors (mRECIST). Patients were categorized as having complete response (CR), partial response (PR), stable disease (SD), or progressive disease (PD). Persistent PD after three or more consecutive standardized TACE sessions was defined as TACE non-response and considered indicative of TACE refractoriness.

Spatial transcriptomics profiling using CosMx™ spatial molecular imaging (SMI)

FFPE tissue sections were analyzed using the 1000-plex CosMx™ Human Universal Cell Characterization RNA Panels on tissue microarray (TMA) slides. The samples underwent standard processing, including deparaffinization, heat-induced epitope retrieval (HIER), enzymatic digestion, and fiducial labeling. Subsequently, multiplexed RNA probes were hybridized in situ, followed by iterative hybridization-imaging cycles on the CosMx™ Spatial Molecular Imager (SMI). For cell segmentation, morphological markers—including CD298/B2M, PanCK, CD45, CD3, and DAPI—were employed. Cell boundaries were defined using an enhanced CellPose algorithm, and transcript-by-cell matrices were generated. To ensure data quality, cells exhibiting elevated negative or system control signals were excluded from downstream analysis.

Cell type annotation

Cell type annotation was conducted through a two-step approach. First, major cellular lineages were classified using well-established marker genes:

• T/NK cells: PTPRC, CD3D, CD3E, NKG7, NCAM1, KLRG1, KLRB1.

• B cells: CD79A, CD19, MS4A1, MZB1, IGHA1, JCHAIN, SDC1.

• Myeloid cells: CD14, CD16, CD68, CD163.

• Endothelial cells: CDH5, VWF, PECAM1.

• Fibroblasts: ACTA2, COL1A1, COL1A2, DCN.

• Hepatocytes: TTR, *SAA1/2*, FGG.

Malignant hepatocytes were distinguished based on spatial origin: cells derived from the tumor core were designated as malignant, whereas those from adjacent non-tumorous tissue were classified as normal.

Subsequently, each major lineage was subsetted and subjected to high-resolution subclustering. Subtype definition was performed using the top 100 differentially expressed genes (DEGs) identified by Seurat’s FindAllMarkers function (logfc.threshold > 0.25, adjusted p-value < 0.05). To ensure cluster purity, putative multiplets—defined as clusters co-expressing markers from multiple lineages—were systematically excluded from downstream analyses.

Spatial proteomics profiling using GeoMx DSP immuno-oncology protein assay

Formalin-fixed, paraffin-embedded (FFPE) tissue sections were analyzed using the GeoMx Digital Spatial Profiler (DSP, Bruker) for high-plex spatial proteomic profiling. Sections were first stained with a panel of fluorescent morphology markers (SMA for stromal cells, CD45 for leukocytes, CK8/18 for epithelial cells, and SYTO13 for nuclear visualization) to facilitate histology-guided region-of-interest (ROI) selection.

The GeoMx Immune-Oncology Proteome Assay (~ 570-plex oligo-tagged antibodies) was then applied to quantify protein expression within spatially resolved ROIs. Following antibody hybridization, UV-mediated photocleavage was performed to release oligo tags exclusively from manually annotated ROIs (tumor core, invasive border, and adjacent normal tissue compartments). The released tags were collected, PCR-amplified, and quantified by Illumina NovaSeq next-generation sequencing. Raw sequencing data were processed using the GeoMx DSP pipeline to generate high-dimensional, spatially resolved proteomic maps, enabling quantitative analysis of protein expression across tissue architectures. Importantly, due to the ROI-based capture design, this technology does not provide single-cell spatial resolution, and the resulting data represent aggregate protein expression within defined tissue regions rather than individual cells.

GeoMx DSP data analysis

Raw sequencing reads were processed through adapter trimming, read merging, and alignment to target probes. Unique molecular identifiers (UMIs) were applied to remove PCR duplicates, ensuring accurate digital quantification. For each sample, protein expression levels were derived by aggregating probe signals, with outliers excluded and the final value defined as the median of retained probes. The limit of quantification was determined as the log-mean of negative control probes plus two standard deviations. Inter-sample variability was estimated using the L2 norm distance. Antibody specificity was assessed by signal-to-noise ratio (SNR), calculated as the target signal over the mean intensity of triplicate IgG controls within corresponding regions; proteins with SNR < 3 were classified as undetectable. Differential protein expression was evaluated using fold-change analysis and Student’s t-test in R.

Quantification of cell type proportions using HALO software

Morphology markers-stained GeoMx tissue slide was scanned at high resolution and analyzed using the HALO^® image analysis platform (Indica Labs). A dedicated multiplex analysis module was employed to segment individual cells based on DAPI nuclear staining and to identify cell types using specific fluorescence markers: CK8/18 for epithelial cells, SMA for stromal cells, and CD45 for immune cells. Signal thresholds for each marker were manually optimized and validated by a board-certified pathologist to ensure accurate cell classification. The relative proportion of each cell type was calculated as the number of positive cells for a given marker divided by the total number of nucleated cells within each ROI. Data were exported from HALO for downstream statistical analysis.

Public data collection

The scRNA-seq datasets consisting of 5 primary untreated and 5 TACE-treated patients, which are from independent, unpaired cohorts derived from different individuals used in this study are downloaded from the public NCBI network for PRJNA793914. The GSE104580 cohort was downloaded from the publicly available Gene Expression Omnibus (GEO) database, which includes 81 TACE responders and 66 non-responders HCC patients. The construction of prognostic models utilized bulk RNA-sequencing data and corresponding clinical data from patients with HCC sourced from The Cancer Genome Atlas (TCGA) database, encompassing 370 tumor samples (TCGA-LIHC). Furthermore, the validation cohort consisted of 240 tumor samples (ICGC-LIRI-JP) obtained from the International Cancer Genome Consortium (ICGC) cohort.

Identification of single cell phenotype associated subpopulation

To quantitatively evaluate the association between individual cells and phenotype-related outcomes, we applied the single-cell Phenotype Associated Score (scPAS) framework, a recently developed computational method that integrates regression modeling with transcriptomic profiles to infer cell-level phenotype associations [18]. Briefly, scPAS constructs a network-regularized sparse regression model that links gene expression features at the single-cell level with clinical phenotypes of interest. In our analysis, the model was trained using survival information from the TCGA_LIHC cohort, enabling the identification of transcriptional programs associated with adverse prognosis. The algorithm assigns each cell a normalized risk score (NRS), which reflects the strength and direction of association between its gene expression profile and the phenotype. Specifically, higher NRS values indicate a stronger positive association with poor survival (high-risk phenotype), whereas lower NRS values indicate association with favorable outcomes. Model construction was performed using the scPAS package with the following parameters: imputation = FALSE, nfeature = 3000, alpha = 0.01, network_class = “SC”, and family = “cox”. These settings enable sparse feature selection while incorporating gene–gene network constraints to improve model robustness.

Pseudo time analysis of scRNA-seq

scTour (v1.0.0), a method based on deep learning model, was used for the trajectory inference and pseudotime analysis [64].

Construction of machine learning model to identify TACE resistant subtypes

Using the XGBoost machine learning algorithm, a model was constructed to predict TACE response based on deconvoluted cell type proportions by performing grid search and k-fold cross-validation with GridSearchCV. The dataset was split into training and test sets at an 4:1 ratio, and the SHAP method was applied to evaluate the contribution of each cell type to the prediction [65].

Pathway activity estimation

Pathway activities were inferred using the R package PROGENy (v1.16.0) with default parameters [21]

Inference of transcription factor regulatory strength

DecoupleR, using the multivariate linear model, was used to calculate the transcription factor activity of CollecTRI transcription factor regulatory networks [22, 66].

Define the spatial niche of spatial transcriptomics data

Spatopic was used for cell neighborhood and domain analysis on LiverCancer1 and LiverCancer2 (ntopics = 15, sigma = 50, region_radius = 400) [67].

Ligand receptor analysis

Intercellular communication analysis was conducted using CellChat (v1.4.0) based on the curated ligand-receptor interaction database (CellChatDB) [68].

Ro/e enrichment analysis

To quantify the enrichment of each cell cluster across different groups, we calculated the ratio of observed to expected cell numbers (Ro/e) as previously described [69]. First, a contingency table of cell clusters by groups was constructed using pooled cells from all patients within each group. A chi-squared test was then applied to estimate the expected distribution. The expected cell number (E_ij) for a given cluster i in group j was calculated based on the marginal totals of the contingency table: E_ij = (R_i × C_j)/N, where R_i is the total number of cells in cluster i, j is the total number of cells in group j, and N is the total number of cells in the table. The Ro/e value was subsequently calculated as Ro/e = Observed/Expected. Ro/e > 1 signifies that a cell state is enriched within a group, while an Ro/e < 1 indicates relative depletion. The resulting values were visualized as heatmaps using the R package pheatmap (v.1.0.13).

FN1 knockdown in PMA-induced THP-1 cells

THP-1 monocytic cells were firstly treated with 100 nM phorbol 12-myristate 13-acetate (PMA) for 48 h to induce macrophage differentiation and then transfected with small interfering RNA (siRNA) targeting FN1 (fibronectin 1) using Lipofectamine RNAiMAX (Thermo Fisher Scientific) according to the manufacturer’s protocol. A non-targeting siRNA served as a negative control. Cells were harvested 48 h post-transfection for downstream analyses. Knockdown efficiency was verified by quantitative real-time PCR and immunoblotting.

Differentiation and RNA sequencing of FN1 knockdown THP-1 cells

Following FN1 knockdown, followed by 24 h of rest in PMA-free medium. Total RNA was extracted using the RNeasy Mini Kit (Qiagen) according to the manufacturer’s instructions. RNA quality was assessed with Agilent Bioanalyzer 2100, and samples with RNA integrity number (RIN) > 8 were used for library preparation. Poly(A)-enriched libraries were constructed with the TruSeq RNA Library Prep Kit (Illumina) and sequenced on the Illumina NovaSeq 6000 platform to generate 150-bp paired-end reads. Differential gene expression analysis was performed after alignment and quantification using standard RNA-seq pipelines.

Quantitative PCR for LGALS3 of FN1 knockdown THP-1 cells

After FN1 knockdown and PMA-induced differentiation of THP-1 cells into macrophages, total RNA was isolated using the RNeasy Mini Kit (Qiagen). Complementary DNA (cDNA) was synthesized from 1 µg of total RNA with the PrimeScript RT Reagent Kit (Takara) following the manufacturer’s protocol. Quantitative PCR was performed using SYBR Green Master Mix (Applied Biosystems) on a QuantStudio 6 Flex Real-Time PCR System. The primer sequences for LGALS3 were: forward 5′-GCCTACCCATCTTCTGGACA-3′ and reverse 5′-GAAGCGTGGGTTAAAGTGGA-3′. GAPDH was used as the internal control. Relative expression levels were calculated using the 2^−ΔΔCt method.

CellScape precise spatial proteomics

Multiplex spatial proteomic profiling was performed using the CellScape™ Precise Spatial Proteomics platform (Bruker), enabling high-plex detection while preserving tissue morphology. FFPE tumor sections were mounted into the whole-slide imaging chamber and iterative cycles of validated antibody staining, high-resolution imaging, and gentle signal removal were conducted using Enhanced Photobleaching in Cyclic Immunofluorescence (EpicIF™) signal removal technology. Single-cell segmentation and phenotyping were performed using the Qupath 0.5.1 to generate spatially resolved cell maps for downstream analysis. The detailed color schemes for each staining cycle of the multiplex experiment are listed in the Supplementary Table.

Supplementary Information

Supplementary Material 1. ^{(15.9MB, tif)}

Supplementary Material 2. ^{(25.7MB, tif)}

Supplementary Material 3. ^{(27.5MB, tif)}

Supplementary Material 4. ^{(28.4MB, tif)}

Supplementary Material 5. ^{(4.1MB, xlsx)}

Abbreviations

TACE

Transarterial chemoembolization

HCC

Hepatocellular carcinoma

TME

Tumor microenvironment

ICIs

Immune checkpoint inhibitors

PFS

Progression-free survival

OS

Overall survival

TAMs

Tumor associated macrophages

scRNA-seq

Single-cell RNA sequencing

mIF

Multiplex immunofluorescence

mIHC

Multiplexed immunohistochemistry

TMAs

Tissue microarrays

FFPE

Formalin-fixed, paraffin-embedded

SMI

Spatial Molecular Imaging

DEGs

Differential expressed genes

Author contributions

F.J., H.C., B.W., X.W., Y.X., J.Z., J.Y., W.Y.,H.L., B.T., H.X., X.L., Y.G., S.W., Q.H., and Y.Z. acquired the data. F.J. and H.C. developed the methodology, performed data analysis and interpretation, and prepared the figures. F.J. drafted the manuscript. J.D., S.Y., and F.J. conceived and designed the study. J.D., S.Y., and Y.G. supervised the study. J.D. and S.Y. were responsible for funding acquisition. All authors read and approved the final version of the manuscript.

Funding

This study was supported by the National Science Foundation of China (82090052, 82090050, and 82090053), Tsinghua University Initiative Scientific Research Program of Precision Medicine (2022ZLA007).

Data availability

The publicly available datasets analyzed in this study were obtained from the original publications. All in-house spatial data are available upon request from the corresponding author.

Declarations

Ethics approval and consent to participate

FFPE tissue specimens were retrospectively collected from four patients with HCC who received TACE treatment followed by surgical resection at the Department of Pathology, Beijing Tsinghua Changgung Hospital. The study protocol was reviewed and approved by the Ethics Committee of Beijing Tsinghua Changgung Hospital, Tsinghua University (Approval No. 25521-0-02), and written informed consent was obtained from all participants.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.