Cell type-resolved analysis identifies LOXL2⁺ fibroblasts as key drivers of malignant stromal remodeling in hepatocellular carcinoma

Abstract

Hepatocellular carcinoma (HCC) typically progresses within a fibrotic, collagen-rich microenvironment where extracellular matrix (ECM) remodeling critically dictates tumor malignancy. However, the cell-type-specific contributions of the lysyl oxidase (LOX) family to this remodeling remain poorly understood. In this study, we conducted an integrative transcriptomic analysis of the LOX gene family (LOX, LOXL1-4) in HCC. Bulk RNA-seq analysis revealed that LOX, LOXL2, and LOXL4 were significantly upregulated in tumors and correlated with advanced pathological stages and poor prognosis. Single-cell RNA-seq (scRNA-seq) analysis delineated distinct expression landscapes: LOX and LOXL4 were enriched in malignant cells, while LOXL2 was predominantly expressed in fibroblasts and endothelial cells. Notably, we identified a discrete LOXL2⁺ fibroblast subset characterized by transcriptomic programs associated with ECM remodeling, contractility, angiogenesis, and hypoxia. A signature derived from LOXL2⁺ fibroblasts was significantly associated with inferior overall survival and enhanced epithelial-mesenchymal transition (EMT) activity in the TCGA-LIHC cohort. Intercellular signaling analysis (CellChat) demonstrated that LOXL2⁺ fibroblasts engage in robust crosstalk with malignant cells via collagen/periostin-integrin signaling axes. Finally, STIP1 was identified through LASSO and random survival forest models as a critical prognostic effector within the LOXL2⁺ fibroblast program. Our findings establish LOXL2⁺ fibroblasts as a pivotal stromal subset driving malignant remodeling and provide potential therapeutic targets for HCC.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-05339-9.

Introduction

Hepatocellular carcinoma (HCC) is the most common primary liver malignancy and represents a significant cause of cancer-related mortality worldwide [1, 2]. It typically arises on a background of chronic liver injury and fibrosis driven by viral hepatitis, alcohol-related liver disease, and metabolic dysfunction-associated steatohepatitis. Despite advances in surgical resection, local ablation, liver transplantation, and systemic therapies including multikinase inhibitors and immune checkpoint blockade, the prognosis of patients with advanced HCC remains unsatisfactory, largely due to late diagnosis, high rates of recurrence, and marked intertumoral and intratumoral heterogeneity [3]. Increasing evidence indicates that the tumor microenvironment (TME), particularly the fibrotic, collagen-rich stroma, is not merely a passive scaffold but an active driver of tumor progression, immune evasion, and therapeutic resistance in HCC [4, 5].

The lysyl oxidase (LOX) gene family comprises five copper-dependent amine oxidases: LOX and its paralogs LOXL1, LOXL2, LOXL3, and LOXL4, which are predominantly secreted enzymes that catalyze the oxidative deamination of lysine residues in collagen and elastin [6]. Family members are encoded on distinct chromosomal loci and share a conserved C-terminal catalytic domain containing the lysine-derived quinone cofactor, copper-binding site, and residues essential for enzymatic activity. In contrast, their N-terminal regions diverge and endow isoform-specific functions [7–9]. LOX and LOXL1 contain pro-peptides whose proteolytic processing, for example by bone morphogenetic protein-1, is required for full activation, whereas LOXL2-4 harbor tandem scavenger receptor cysteine-rich (SRCR) domains [10]. By generating reactive aldehydes that mediate covalent cross-linking of extracellular matrix (ECM) components, LOX family enzymes critically regulate ECM architecture, tissue stiffness, and mechanotransduction. Aberrant upregulation of LOX family activity has been implicated in pathological fibrosis of multiple organs, including the liver, and in the initiation and progression of diverse malignancies. In particular, LOXL2 has been extensively linked to desmoplasia, epithelial-mesenchymal transition, and metastasis across tumor types, while LOXL3 and LOXL4 have also been reported to support tumor growth and dissemination in several cancers, including HCC [11–13].

However, in HCC, the biology of the LOX family remains incompletely defined at both the molecular and cellular levels. It is unclear which LOX family members are consistently dysregulated in human HCC, how their expression relates to clinicopathological features and patient outcome, and which cell populations within the TME are the principal sources and responders to LOX-mediated signaling. In this study, we performed an integrative analysis of LOX family genes in HCC by bulk RNA-seq and scRNA-seq. We first characterized the expression and clinicopathological associations of LOX, LOXL1, LOXL2, LOXL3, and LOXL4 in HCC and adjacent normal liver. We then delineated their cell type-specific expression patterns within the TME of HCC. We identified a LOXL2-positive fibroblast population with a transcriptional program linked to epithelial-mesenchymal transition, angiogenesis, and metabolic reprogramming. Finally, we evaluated the prognostic relevance and intercellular signaling networks of LOXL2-expressing fibroblasts in relation to malignant cells, thereby providing a comprehensive framework for understanding LOX family-mediated stromal remodeling and its contribution to HCC progression.

Methods

Clinical samples and experimental validation

A total of 30 pairs of hepatocellular carcinoma (HCC) and adjacent non-tumor tissues were obtained from patients who underwent surgical resection at Jinhua Municipal Central Hospital. All procedures were approved by the Ethics Committee of Jinhua Municipal Central Hospital (Approval No.202663).

Data acquisition and preprocessing

Transcriptome data (TPM format), raw counts, and corresponding clinical annotations were obtained from The Cancer Genome Atlas (TCGA) using the R package TCGAbiolinks (v2.26.0) [14]. After removing samples with incomplete clinical information or zero survival time, 369 tumor samples and 160 adjacent normal tissues were retained for downstream analyses. Genes with > 50% missing values or expressed in fewer than 50% of samples were excluded. TPM values were log₂-transformed as log₂(TPM + 1) for visualization and statistical analyses.

To validate protein-level expression of LOX family members, we retrieved mass spectrometry-based proteomic profiles of HCC from the CPTAC data portal. Immunohistochemistry (IHC) images of LOXL2 in normal liver and HCC were downloaded from the Human Protein Atlas (HPA).

Gene mutation annotation and domain mapping for LOX family genes were obtained from cBioPortal and visualized using maftools (v2.16.0) [15].

Bulk RNA-seq differential expression and clinicopathological analyses

Differential expression between tumor and normal liver tissues in TCGA-LIHC was assessed using DESeq2. Where applicable, p-values were adjusted for multiple testing using the Benjamini-Hochberg method. Clinicopathological associations across age groups or pathological stages were evaluated using Wilcoxon test depending on data distribution. Protein-level differences in CPTAC were analyzed using Moderated t-test.

Cell-type expression profiling using TISCH

To assess the cellular distribution of LOX family genes within the tumor microenvironment, we queried multiple single-cell HCC datasets from the TISCH2 database. Expression matrices were downloaded directly from the online portal, and normalized cell-type-specific expression was visualized across curated cell lineages, including malignant cells, fibroblasts, endothelial cells, myeloid cells, lymphocytes, and mast cells.

scRNA-seq data processing and fibroblast subsetting

The scRNA-seq dataset GSE149614, consisting of primary hepatocellular carcinoma samples, was obtained from the GEO database and processed using Seurat (v5.1.0) following standard analytical procedures [16, 17]. After quality control filtering to remove cells expressing fewer than 200 genes, more than 6000 genes, or exceeding 20% mitochondrial transcript content. The DoubletFinder R package removed the doublet. The data were log-normalized and scaled using NormalizeData and ScaleData. Highly variable genes were identified with FindVariableFeatures, and dimensionality reduction was performed through principal component analysis followed by UMAP embedding. Cell clustering was conducted using a graph-based approach implemented in FindClusters. Cell-type annotation was conducted by manual refinement based on known marker genes. Fibroblasts were isolated from the integrated object using canonical stromal markers (COL1A1, COL1A2, COL3A1, LUM, DCN), and LOXL2-positive fibroblasts were defined as those with LOXL2 expression levels in the upper quartile of the fibroblast population.

Fibroblast subtype comparison with published references

To improve biological interpretation of fibroblast subclusters, we compared them with published fibroblast reference signatures using GSVA reference mapping. Signature enrichment scores were computed for each fibroblast cluster, and cluster annotation was refined according to the highest similarity to established fibroblast subtype programs.

Differential expression analysis in fibroblast subsets

Differentially expressed genes (DEGs) between LOXL2-positive fibroblasts and LOXL2-negative fibroblasts were computed using FindMarkers with the Wilcoxon rank-sum test. Genes with |log2 fold change| ≥ 0.5 and adjusted p < 0.05 were considered significantly dysregulated. Volcano plots were generated using EnhancedVolcano [18].

Gene set enrichment analysis

Functional enrichment of DEGs was assessed using clusterProfiler (v4.10.0) [19, 20]. Both Hallmark and GO Biological Process gene sets (MSigDB v2024.1) were interrogated. GSEA significance was defined as: |NES| > 1 and nominal p < 0.05.

Cell-cell communication analysis

To evaluate ligand-receptor interactions between fibroblast subsets and malignant cells, we used CellChat (v1.6.0) [21]. Analyses included: computation of outgoing and incoming signaling strengths, identification of enriched pathways, visualization of interaction networks, and heatmaps and bubble plots of key ligand-receptor pairs. Comparisons of LOXL2-positive vs. LOXL2-negative fibroblasts were performed using differential communication analyses.

Bulk RNA-seq correlation and signature scoring

To quantify EMT activity in TCGA, we calculated EMT pathway scores using GSVA (v1.50.5) with the HALLMARK_EMT gene set. Correlation between LOXL2 expression (bulk level) or LOXL2-positive fibroblast signature (single-cell-derived) and EMT scores was evaluated using Spearman correlation. A LOXL2-positive fibroblast signature score was computed using AddModuleScore in Seurat and projected onto TCGA samples via ssGSEA. Detailed information for all independent datasets, including sample sizes and platform details, is summarized in Supplementary Table S2.

Survival analysis

Overall survival (OS) was assessed in TCGA-LIHC using survival (v3.5-7) and survminer (v0.4.9). Patients were dichotomized into high- and low-signature groups based on median LOXL2-positive fibroblast score or LOXL2 expression. Kaplan-Meier curves were generated, and significance was assessed using the log-rank test.

LASSO regression and random survival forest

Least absolute shrinkage and selection operator (LASSO) Cox regression was performed with the glmnet R package (v4.1-8), using 10-fold cross-validation to determine the optimal penalty parameter λ and retain genes with non-zero coefficients as prognostic candidates [22]. Genes with non-zero coefficients at lambda.min were retained as prognostic candidates. These retained genes were then input into a random survival forest model using randomForestSRC (randomForestSRC, v3.2.3; 1,000 trees, log-rank splitting rule) to compute permutation-based variable importance [23].

Experimental validationRT-qPCR

Total RNA was extracted using TRIzol reagent (Invitrogen, 15596018CN). cDNA was synthesized and quantified using SYBR Green Master Mix (Thermo Fisher, A46110). Relative expression was calculated via the2^−ΔΔCt method, with GAPDH as an internal control. Primer sequences and amplification conditions are provided in Supplementary Table 1.

Immunohistochemistry (IHC): Formalin-fixed paraffin-embedded (FFPE) sections were incubated with anti-LOXL2 (1:4000, Proteintech, 67139-1-Ig) primary antibodies overnight at 4 °C. Staining intensity and percentage of positive cells were scored by two independent pathologists.

Statistical analyses

Unless otherwise specified, statistical analyses were performed in R (v4.4.1). Data visualization was carried out using ggplot2 (v3.5.1), ComplexHeatmap (v2.18.0), and base R graphics. All tests were two-sided, and a p-value < 0.05 was considered statistically significant unless stated otherwise.

Results

Lox family dysregulation in HCC

Given the established role of the LOX family in extracellular matrix remodeling and tumor progression, we first characterized their expression pattern in HCC. Analysis of TCGA-LIHC RNA-seq data showed that LOX, LOXL2 and LOXL4 transcripts were upregulated in tumor tissues compared with normal liver, whereas LOXL1 and LOXL3 displayed no apparent difference between the two groups (Fig. 1A). To determine whether these changes were reflected at the protein level, we interrogated CPTAC proteomic data for LIHC, which revealed increased LOXL2 protein abundance in primary tumors relative to normal liver, with more modest variation for the other family members (Fig. 1B). Consistently, immunohistochemical staining from the Human Protein Atlas demonstrated stronger LOXL2 staining in HCC specimens than in non-tumor liver tissues (Fig. 1C). Then, we explored the association between LOX family expression and clinicopathologic features in TCGA-LIHC and found that LOX and LOXL1 expression tended to be higher in younger patients. At different stages, LOX and LOXL2 were more abundant in advanced-stage tumors than in early-stage tumors (Fig. 1D). Survival analyses further indicated that higher expression of LOX, LOXL2, LOXL3, and LOXL4 was associated with poorer overall survival, whereas LOXL1 expression showed no clear prognostic separation between high- and low-expression groups (Fig. 1E).

Fig. 1.

Expression, clinicopathologic associations and survival of LOX family genes in HCC. A Boxplots showing mRNA expression of LOX, LOXL1, LOXL2, LOXL3 and LOXL4 in tumor and adjacent normal liver tissues from the TCGA-LIHC cohort. Expression values are displayed as log₂(TPM + 1). B Protein expression of LOX family members in normal liver and primary HCC samples from the CPTAC LIHC proteomic dataset, presented as Z-scores. C Representative immunohistochemistry images from the Human Protein Atlas illustrating LOXL2 staining in HCC and non-tumor liver tissues. D Distribution of expression of LOX family genes across different age groups and pathological stages in TCGA-LIHC. E Kaplan-Meier overall survival curves for TCGA-LIHC patients stratified into high- and low-expression groups for each LOX family gene (LOX, LOXL1-4); group separation is based on the median expression level, and p-values are derived from log-rank tests. For all panels, statistical significance is denoted as follows: ns, not significant; *P < 0.05; P < 0.01; ***P < 0.001

Genomic alterations of LOX family genes in HCC

To characterize the mutational landscape of the LOX family in HCC, we analyzed somatic variants from the cBioPortal database and mapped them onto the corresponding protein domains. Across all five genes, mutations were relatively infrequent and predominantly scattered throughout the coding regions without evidence of recurrent hotspots (Fig. 2A–E). LOX displayed variants distributed along the protein sequence, with occasional changes located near the lysyl-oxidase catalytic domain. LOXL1 showed a similarly sparse pattern, with mutations largely confined to non-repetitive regions of the protein. LOXL2 exhibited a broader distribution of variants across its multiple SRCR domains and the C-terminal catalytic region, although no single site appeared repeatedly altered. LOXL3 and LOXL4 harbored mutations across both the SRCR modules and the catalytic domain, again with low patient frequency for each event. Overall, these data suggest that while LOX family genes can be somatically mutated in HCC, such alterations occur at low rates and lack defined hotspot residues, implying that transcriptional and post-transcriptional dysregulation may contribute more substantially to their aberrant activity in HCC.

Fig. 2.

Somatic mutation and domain architecture of LOX family proteins in HCC. A–E Lollipop plots depicting the distribution of non-synonymous somatic mutations in LOX (A), LOXL1 (B), LOXL2 (C), LOXL3 (D) and LOXL4 (E) in TCGA-LIHC. Vertical sticks indicate individual mutated residues and the number of patients carrying each variant. Colored horizontal blocks represent annotated protein domains, including SRCR modules and the C-terminal lysyl oxidase catalytic domain. Amino acid positions are shown on the x-axis and patient counts on the y-axis. Specifically, green indicates missense mutations, yellow indicates splice mutations, and gray indicates samples truncating mutation types

Cellular distribution of LOX family expression in the TME of HCC

To determine which cellular compartments drive the dysregulated LOX family expression in HCC, we examined their expression patterns across annotated cell populations in multiple scRNA-seq datasets from TISCH database. LOX transcript abundance was highest in malignant epithelial cells and fibroblasts, indicating that both tumor cells and stromal remodeling populations contribute to its upregulation (Fig. 3A). LOXL1 expression was primarily restricted to fibroblasts, consistent with its established role in extracellular matrix organization (Fig. 3B). LOXL2 showed a broader but still stromal-skewed distribution, with predominant expression in fibroblasts and endothelial cells (Fig. 3C). LOXL3 was expressed across nearly all major immune and stromal compartments without strong lineage specificity (Fig. 3D). LOXL4 was enriched in malignant cells and displayed detectable expression in epithelial, endothelial, fibroblast, monocyte/macrophage, and mast cell subsets (Fig. 3E), suggesting a transcriptional program shared across multiple cell types.

Fig. 3.

Cell-type-specific expression patterns of LOX family genes in the TME of HCC. A–D Heatmaps showing normalized expression of LOX (A), LOXL1 (B), LOXL2 (C), LOXL3 (D) and LOXL4. E across major cell lineages in multiple LIHC scRNA-seq datasets curated in the TISCH2 database. F UMAP plot of the GSE149614 scRNA-seq dataset, colored by major cell types. G UMAP feature plots displaying the distribution and relative expression levels of LOX, LOXL1, LOXL2, LOXL3 and LOXL4 in GSE149614

We next validated these observations using an independent single-cell RNA-seq dataset (GSE149614). UMAP showed major cellular compartments of the TME of HCC, including malignant cells, hepatocytes, fibroblasts, endothelial cells, myeloid cells, and lymphocytes (Fig. 3F). Mapping LOX family transcripts onto the UMAP space recapitulated the trends derived from the TISCH database (Fig. 3G). Together, these results identify distinct cell-type-specific expression patterns across LOX family members and highlight fibroblasts and malignant cells as primary transcriptional sources in HCC.

LOXL2 is upregulated in hepatocellular carcinoma and correlates with advanced tumor stage

To determine LOXL2 expression in relation to the severity of hepatocellular carcinoma (HCC), we first compared its expression in tumor tissues and paired adjacent non-tumor tissues. qPCR revealed that LOXL2 mRNA expression was significantly elevated in HCC tissues compared with adjacent non-tumor counterparts (Fig. 4A). Moreover, qPCR analysis also demonstrated LOXL2 transcription was positively associated with HCC stage. Using the Barcelona Clinic Liver Cancer (BCLC) staging system, we found that LOXL2 mRNA levels gradually increased from early-stage (BCLC 0) to advanced-stage disease (BCLC ≥ B) (Fig. 4B). Consistently, the tendency was similar when we used a China Liver Cancer (CNLC) staging system (Fig. 4C). These findings suggest a strong correlation between elevated LOXL2 transcription and HCC progression. We next validated the association at the protein level. IHC (Immunohistochemistry) staining revealed that the LOXL2 protein expression in HCC tumor tissue was significantly higher than in adjacent tissue. Moreover, consistent with the mRNA data, LOXL2 protein expression progressively increased with advancing BCLC stage, with the highest levels observed in BCLC ≥ B tumors (Fig. 4E).

Fig. 4.

LOXL2 is upregulated in hepatocellular carcinoma and correlates with tumor progression. A Relative mRNA expression of LOXL2 in paired HCC tumor tissues (n = 30) and adjacent non-tumor tissues (n = 30) determined by RT-qPCR. B LOXL2 mRNA expression stratified by BCLC stage (0, A, ≥B), demonstrating a progressive increase of LOXL2 expression with advancing tumor stage. C LOXL2 mRNA expression according to CNLC stage (Ia, Ib, ≥IIa), further confirming its association with advanced disease stages. D Representative IHC staining of LOXL2 in adjacent non-tumor and tumor tissues (left), with quantification of LOXL2-positive cells per 40× field (right), showing significantly higher LOXL2 protein expression in tumor tissues. Scale bars: 100 μm (10×magnification) and 25 μm (40× magnification). E Representative IHC staining of LOXL2 in HCC tissues across different BCLC stages (0, A, ≥B) (left), and corresponding quantification of LOXL2-positive cells per 40× field (right), indicating increased LOXL2 protein expression with tumor progression. Data are presented as mean SD. Statistical significance was determined by paired t-test (A, D) or one-way ANOVA followed by Tukey’s post-hoc test (B, C, E). ***p<0.001, **** p<0.0001

Collectively, these data demonstrate that LOXL2 is significantly upregulated in HCC at both the mRNA and protein levels and is positively associated with HCC tumor stage, highlighting its potential role in HCC progression and its clinical relevance as a marker of aggressive disease.

LOXL2⁺ fibroblasts define a pro-tumorigenic stromal program associated with poor prognosis

Given the observed association between upregulated LOXL2 expression and unfavorable clinical outcome, together with its preferential enrichment in fibroblasts, we sought to determine whether LOXL2-expressing fibroblasts constitute a stromal state linked to heightened tumor aggressiveness and to elucidate their potential functional role within the TME. We performed a focused analysis on the fibroblast compartment of this HCC scRNA-seq dataset. After dimension reduction and unsupervised clustering, 7 clusters were identified (Fig. 5A). We find LOXL2 highly expressed in C0 compared with other clusters. (Fig. 5A, B). Comparison with published fibroblast reference signatures further suggested that this cluster most closely resembled myofibroblastic fibroblasts (Supplementary Fig. 1A).

Fig. 5.

Single-cell characterization of LOXL2⁺ fibroblasts and their transcriptional and communication programs. A UMAP embedding of fibroblasts from GSE149614, colored by unsupervised clusters (C0-C6). B UMAP feature plot showing LOXL2 expression within the fibroblast compartment, highlighting the LOXL2⁺ fibroblast cluster. C Heatmap of selected marker genes across fibroblast clusters, illustrating transcriptional features of fibroblast subclusters. D Kaplan-Meier overall survival curves for TCGA-LIHC patients stratified by a LOXL2⁺ fibroblast signature score (high vs. low; median cut-off). Survival differences are assessed by the log-rank test. E Volcano plot of differentially expressed genes between LOXL2⁺ fibroblasts and LOXL2⁻ fibroblasts derived from scRNA-seq data; significantly up- and downregulated genes are highlighted. F Gene set enrichment analysis (GSEA) of Hallmark pathways based on differential expression between LOXL2⁺ and LOXL2⁻ fibroblasts, shown as ranked t-scores. G Bubble plot from CellChat analysis depicting selected ligand-receptor interactions between LOXL2⁺ fibroblasts and malignant cells; circle size reflects statistical significance and color denotes communication probability. H Heatmap summarizing the number of inferred ligand-receptor interactions from LOXL2⁺ and LOXL2⁻ fibroblasts to hepatocytes. I, J Scatter plots showing the correlation between EMT pathway scores and either bulk LOXL2 expression (I) or LOXL2⁺ fibroblast signature scores (J) in TCGA-LIHC; regression lines and Spearman correlation coefficients are displayed

These LOXL2⁺ fibroblasts displayed a transcriptional profile markedly different from other fibroblast states. (Fig. 5C). To assess their clinical relevance, we estimated the abundance of LOXL2⁺ fibroblasts in TCGA-LIHC and examined their association with patient outcome. Tumors with higher LOXL2⁺ fibroblast signatures showed a trend toward inferior overall survival compared with those with lower signatures (Fig. 5D; log-rank p = 0.062), indicating that this stromal program reflects a more aggressive tumor.

Differential expression analysis further confirmed the malignant-related phenotype of LOXL2⁺ fibroblasts. Compared with LOXL2⁻ fibroblasts, LOXL2⁺ fibroblasts upregulated a wide array of pro-tumorigenic genes, including collagens, ECM cross-linking enzymes, contractile cytoskeletal components, and factors implicated in matrix stiffening and angiogenesis (Fig. 5E). Functional enrichment analyses highlighted pathways involved in epithelial-mesenchymal transition, angiogenesis, glycolysis, inflammatory responses, hypoxia adaptation, and multiple oncogenic signaling modules (Fig. 5F), indicating that LOXL2⁺ fibroblasts are tightly coupled to a tumor-supportive stromal state.

Next, we profiled cell-cell communication to determine how LOXL2⁺ fibroblasts interact with malignant epithelial cells. CellChat analysis showed that LOXL2⁺ fibroblasts participated in a greater number and a broader diversity of inferred ligand-receptor interactions with tumor cells than other fibroblast subsets. However, interaction strength analysis indicated that this did not translate into uniformly stronger global communication. In some contexts, other fibroblast subsets exhibited comparable or even higher overall signaling strength. Therefore, our data support the notion that LOXL2⁺ fibroblasts possess a distinct communication repertoire, particularly enriched in ECM- and adhesion-related signaling pathways, rather than simply stronger overall reciprocal signaling (Fig. 5G and Supplementary Fig. 1B). Examination of ligand-receptor pairs revealed that fibroblast-to-tumor signaling was dominated by ECM-receptor interactions (such as COL1A1/2-integrin axes) and matricellular cues (for example, POSTN-ITGAVB5), pathways known to promote tumor invasion, extracellular matrix remodeling, and metabolic reprogramming (Fig. 5H).

Given the strong enrichment of EMT signatures in LOXL2⁺ fibroblasts, we finally assessed their relationship with EMT in bulk samples from the TCGA database. LOXL2 expression in TCGA-LIHC correlated strongly with EMT program activity, and the LOXL2⁺ fibroblast signature similarly showed a positive association with EMT scores (Fig. 5I, J). These findings support a model in which LOXL2-expressing fibroblasts foster a pro-invasive TME conducive to malignant progression.

Identification of STIP1 as a prognostic LOXL2⁺ fibroblast-related target

To pinpoint key prognostic effectors within the LOXL2⁺ fibroblast transcriptional program, we first applied LASSO Cox regression to genes enriched in LOXL2⁺ fibroblasts, which reduce the initial feature set from 26 genes to 10 genes with non-zero coefficients with stable survival associations (Fig. 6A). These candidates were then entered into a random survival forest model to quantify their relative contribution to overall survival. The variable importance ranking highlighted STIP1 as the top feature among LOXL2⁺ fibroblast-related genes, followed by several metabolism- and stress-response-associated molecules (Fig. 6B). On this basis, we examined the prognostic impact of STIP1 alone. We found that patients with high STIP1 expression had markedly shorter overall survival than those with low expression (Fig. 6C), supporting STIP1 as a core LOXL2⁺ fibroblast-associated target with adverse outcome in HCC.

Fig. 6.

Identification of prognostic LOXL2⁺ fibroblast-related targets in HCC. A Ten-fold cross-validation curve of LASSO Cox regression applied to LOXL2⁺ fibroblast-enriched genes in TCGA-LIHC. The x-axis shows log(λ), the y-axis shows partial likelihood deviance, and numbers above the curve indicate the number of non-zero coefficients at each penalty value. B Variable importance plot from a random survival forest model built using LASSO-selected genes, ranking LOXL2⁺ fibroblast-related genes according to their contribution to overall survival. C Kaplan-Meier overall survival curves for TCGA-LIHC patients stratified by STIP1 expression (high vs. low; median cut-off)

Discussion

In this study, we systematically characterized the LOX family in HCC using integrated bulk transcriptomic, proteomic, mutational, and single-cell transcriptomic analyses. While LOXL2 overexpression and its association with extracellular matrix remodeling have been reported previously in HCC, by comparing the entire LOX family in a unified framework; second, by resolving their cell-type-specific expression patterns at single-cell resolution; and third, by identifying a LOXL2-enriched fibroblast program associated with stromal activation, EMT-related features, and adverse clinical outcome. Studies across multiple tumour types have established that LOX, LOXL1, LOXL2, LOXL3, and LOXL4 are secreted copper-dependent amine oxidases that catalyse the cross-linking of collagen and elastin [24, 25]. This activity increases extracellular matrix stiffness, facilitates EMT and invasive behaviour, and in several settings contributes to the formation of immunosuppressive and metastatic niches. However, these five enzymes are not functionally equivalent: LOX and LOXL1 are classical collagen and elastin cross-linkers that are strongly induced by hypoxia and TGF-β, LOXL2 is a potent driver of desmoplasia and EMT, LOXL3 stabilizes EMT transcription factors and supports survival, and LOXL4 has been linked to exosome-mediated metastasis and immune evasion [26–29]. Our systematic analysis in HCC confirms this functional heterogeneity at the transcriptional and protein levels. We show that LOX, LOXL2, and LOXL4, but not LOXL1 or LOXL3, are consistently upregulated in tumor tissue relative to non-tumor liver, and that LOXL2 in particular is concordantly elevated in proteomic and immunohistochemical datasets. The association of LOX and LOXL2 with advanced pathological stage further aligns with their known roles in ECM stiffening, mechano-transduction, and metastatic spread [30]. At the same time, somatic mutations in LOX family genes are rare and non-recurrent in HCC, reinforcing the view that their contribution to liver carcinogenesis is driven by transcriptional and microenvironmental regulation rather than by classical oncogenic mutations [24, 31]. By resolving expression at single-cell resolution, we extend prior bulk observations and demonstrate that different cellular compartments preferentially deploy distinct LOX family members: LOX and LOXL4 are concentrated in malignant cells (with LOX additionally expressed by fibroblasts), LOXL1 is confined mainly to fibroblasts, LOXL2 is enriched in fibroblasts and endothelial cells, and LOXL3 is broadly expressed at lower levels across multiple stromal and immune lineages [32, 33]. These patterns suggest that, within HCC, LOX and LOXL4 primarily support tumor cell-intrinsic exploitation of the ECM and immune niche, LOXL1 marks fibroblast-mediated matrix remodeling without a clearly malignant program, LOXL3 may reflect more general stress and EMT-supportive functions, and LOXL2 sits at the intersection of stromal remodeling and vascular niches [34]. This family-wide, cell-type-resolved mapping represents a conceptual advance over previous work that typically focused on single LOX paralogs or could not disentangle tumor and stromal sources.

Within this framework, our data highlight LOXL2⁺ fibroblast as a key stromal node linking LOX family biology to malignant behavior in HCC. LOXL2 has been widely implicated in desmoplasia, EMT induction, pre-metastatic niche formation, and immune exclusion in breast, lung, and gastrointestinal cancers, and several studies in gastric and colorectal tumors have pointed to LOXL2 as a defining component of cancer-associated fibroblast signatures associated with poor prognosis. In HCC, however, LOXL2 has been mainly studied at the bulk level or in tumor cells, and the specific contribution of LOXL2-expressing fibroblasts has remained uncertain [35]. By focusing on the fibroblast compartment in single-cell data, we identify a discrete LOXL2⁺ fibroblast subset with a transcriptional program distinct from other fibroblast states. This subset is enriched for collagen isoforms, cross-linking enzymes, contractile cytoskeletal components, and matricellular proteins, consistent with a highly activated, matrix-remodeling CAF phenotype. Pathway analysis further reveals coordinated upregulation of EMT, angiogenesis, glycolysis, hypoxia responses and diverse oncogenic signaling pathways, indicating that LOXL2 + fibroblasts encapsulate a malignant stromal program rather than a generic fibrotic reaction. Importantly, a LOXL2⁺ fibroblasts-derived signature associates with worse overall survival and higher EMT activity in TCGA-LIHC, linking this stromal state to clinically aggressive disease. Cell-cell communication analysis refines this picture by showing that LOXL2⁺ fibroblasts maintain more intensive and diverse ligand-receptor interactions with malignant cells than other fibroblast subsets, prominently involving collagen-integrin axes, periostin-integrin interactions and midkine-receptor signaling. These pathways are well known to transmit mechanical and biochemical cues that foster tumor cell motility, plasticity, survival and metabolic adaptation [36]. When viewed in light of prior work demonstrating that LOXL2 stiffens the matrix, stabilizes EMT transcription factors and conditions distant organs for metastasis, our findings support a model in which LOXL2⁺ fibroblasts act as a central stromal hub that couples ECM cross-linking to EMT activation and pro-invasive signaling in HCC [37, 38]. In this model, LOXL2 distinguishes itself from other LOX paralogs by its dual capacity to drive a desmoplastic CAF state and to engage malignant cells through a dense ligand-receptor network. In contrast, LOX, LOXL1, LOXL3 and LOXL4 play more restricted or context-dependent roles in this particular fibroblast-tumor axis.

From a translational perspective, these results position LOXL2 and the LOXL2⁺ fibroblast program as promising biomarkers and therapeutic entry points in HCC, while also illustrating the value of a family-wide, multi-omic approach to matrix-modifying enzymes [39]. The robust and multi-level induction of LOXL2, its preferential localization to fibroblasts and endothelial cells in the TME, and the association of LOXL2⁺ fibroblast signatures with poor outcome and heightened EMT suggest that measuring LOXL2 expression or LOXL2⁺ fibroblast abundance could aid in risk stratification and in identifying patients with highly desmoplastic, EMT-enriched tumors [40]. Therapeutically, our data provide mechanistic support for targeting LOXL2 in HCC, not only at the level of tumor cells but specifically at the fibroblast-tumor interface [37]. In principle, pharmacologic LOXL2 inhibition or disruption of key LOXL2⁺ fibroblast-tumor ligand-receptor axes could be combined with immune checkpoint blockade or anti-angiogenic therapies to overcome stromal barriers and immunosuppressive niches [38, 39]. Several limitations should be acknowledged. First, the study is primarily based on retrospective public datasets and is therefore correlative. Functional perturbation experiments will be required to determine whether LOXL2-positive fibroblasts actively contribute to tumor progression [33, 36]. Second, the analyzed single-cell cohorts represent a limited number of patients and may not capture the full biological diversity of HCC. Third, fibroblast subtype annotation depends in part on clustering strategy and available reference signatures. Fourth, the LOXL2-positive fibroblast signal in bulk RNA-seq was represented by a signature enrichment score, which is an indirect surrogate rather than a direct measurement of cell abundance. Moreover, while our communication analyses and EMT correlations are consistent with close physical and functional coupling between LOXL2⁺ fibroblasts and malignant cells, spatial transcriptomics and multiplex imaging will be needed to confirm their spatial organization and to define how this stromal program intersects with immune cell recruitment and function [40]. Future studies that integrate spatially resolved omics, functional genomics and pharmacologic inhibition in preclinical HCC models will be essential to validate LOXL2⁺ fibroblasts as a therapeutic target and to determine how LOX, LOXL1, LOXL3 and LOXL4 cooperate or compete with LOXL2 under different microenvironmental conditions. Despite these caveats, our work provides a family-wide, cell-type-resolved atlas of LOX enzymes in HCC. It identifies LOXL2 + fibroblasts as a critical stromal population that links extracellular matrix remodeling to EMT, pro-invasive signaling, and poor clinical outcome.

Conclusion

In summary, our integrative analyses reveal distinct cell-type-specific expression patterns of LOX family members in hepatocellular carcinoma. Among them, LOXL2 showed a predominantly stromal distribution and was enriched in a fibroblast subset characterized by extracellular matrix remodeling, contractility, angiogenesis- and hypoxia-related programs, and association with EMT. A bulk-derived LOXL2-positive fibroblast signature was linked to unfavorable clinical outcome, supporting the clinical relevance of this stromal program in HCC. These findings refine the current understanding of LOXL2 in HCC by highlighting its stromal cellular context and suggest that LOXL2-positive fibroblasts may represent a biomarker-associated and therapeutically relevant component of malignant stromal remodeling.

Abbreviations

HCC

Hepatocellular carcinoma

LOX

Lysyl oxidase

ECM

Extracellular matrix

RNA-seq

RNA sequencing

scRNA-seq

Single-cell RNA sequencing

TCGA

The Cancer Genome Atlas

LIHC

Liver hepatocellular carcinoma

EMT

Epithelial-mesenchymal transition

LASSO

Least absolute shrinkage and selection operator

TME

Tumor microenvironment

CPTAC

Clinical Proteomic Tumor Analysis Consortium

IHC

Immunohistochemistry

HPA

Human Protein Atlas

SRCR

Scavenger receptor cysteine-rich

TPM

Transcripts per million

GEO

Gene Expression Omnibus

UMAP

Uniform Manifold Approximation and Projection

DEGs

Differentially expressed genes

GSEA

Gene set enrichment analysis

NES

Normalized enrichment score

GO

Gene Ontology

GSVA

Gene set variation analysis

ssGSEA

Single-sample gene set enrichment analysis

OS

Overall survival

FFPE

Formalin-fixed paraffin-embedded

RT-qPCR

Quantitative reverse transcription polymerase chain reaction

GAPDH

Glyceraldehyde-3-phosphate dehydrogenase

BCLC

Barcelona Clinic Liver Cancer

CNLC

China Liver Cancer

ANOVA

Analysis of variance

CAF

Cancer-associated fibroblast

Author contributions

Haijun Chen: Data curation (equal); investigation (equal); project administration (equal); writing – original draft (lead). Lujian Zhu: Data curation (equal); formal analysis (equal); writing – original draft (equal). Junmei Lin: Data curation; project administration. Hongjun Hua: Data curation, writing – original draft. Xuxing Ye: Experimental validation, writing – original draft. Chao Yang: Writing - original draft; revision. Xiaobo Wang: Data curation; formal analysis; investigation; project administration (lead); writing – original draft (lead).

Funding

This study was supported by the Jinhua Science and Technology Plan Project (Social Development Project) (Grant Nos. 2023-3-097, 2024-3-043, and 2025-4-064) and the Special Fund for Basic Research of Jinhua Central Hospital (Grant Nos. JY2023-5-02 and JY2023-6-07).

Data availability

The datasets analyzed during the current study are available in the following public repositories: GEO: scRNA-seq dataset GSE149614 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE149614).TCGA-LIHC: Transcriptome and clinical data accessed via the GDC Data Portal (https://portal.gdc.cancer.gov/projects/TCGA-LIHC).CPTAC & UALCAN: CPTAC liver cancer (tumor vs. normal) proteomic comparisons were performed using the processed CPTAC protein abundance data as presented in the UALCAN database ( [http://ualcan.path.uab.edu/](http:/ualcan.path.uab.edu) ). These mass spectrometry data were utilized for secondary analysis in this study and were not generated by the authors.Human Protein Atlas (HPA): LOXL2 immunohistochemistry images (https://www.proteinatlas.org/).cBioPortal: Mutation and domain annotations for LOX family genes (https://www.cbioportal.org/).TISCH2: Single-cell HCC expression matrices (http://tisch.comp-genomics.org/).

Declarations

Ethics approval and consent to participate

All data used in this study were obtained from publicly available databases (GEO, TCGA), which have undergone independent ethical review and approval by their respective institutional review boards prior to public release. The datasets used in this analysis (TCGA Level 3 data) are standardized and de-identified, thus no additional ethical approval was required for this study. All data usage complied strictly with the data access policies and sharing agreements of GEO (https://www.ncbi.nlm.nih.gov/geo/), TCGA (https://www.cancer.gov/tcga). For the clinical samples analyzed in this study, ethical approval was obtained from the Ethics Committee of Jinhua Municipal Central Hospital (Approval No.202663), and the research was conducted in accordance with the Declaration of Helsinki. The requirement for written informed consent was waived by the Ethics Committee of Jinhua Municipal Central Hospital due to the retrospective nature of the study and the use of anonymized clinical data.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.