Imaging mass cytometry reveals novel cellular neighborhoods for predicting the prognosis of nonalcoholic fatty liver disease-associated hepatocellular carcinoma

Abstract

Objective

We aimed to examine the panoramic cellular composition and spatial structure of clinical samples from nonalcoholic fatty liver disease-associated hepatocellular carcinoma (NAFLD-HCC), assess the landscape of the tumor microenvironment (TME) and identify potential clinical biomarkers for prognosis.

Methods

Imaging mass cytometry (IMC) technology was used to analyze biomarker expression and spatial distribution in samples from 56 patients with NAFLD-HCC. Thirty-seven distinct metal-conjugated antibodies were designed to label the tumor, border and adjacent tissue regions, producing 367 highly multiplexed histological images at single-cell resolution. By combining biomarker expression levels with relative cell positional relationships, we defined functional units as tightly interacting cellular neighborhoods (CNs). Univariate and multivariate Cox regression models were employed to identify independent risk factors for prognosis assessment in patients with NAFLD-HCC on the basis of the proportion of CNs and other conventional clinical indicators.

Results

Eighteen clusters of cells and 9 CNs were defined. Heterogeneity in the TME and CN composition was observed not only across various regions within the same individual but also among comparable regions in different patients. Overall survival data revealed that patients with high enrichment of GPC3⁺/YAP⁺ cancer cell-associated CNs had poor clinical prognoses.

Conclusion

Our study provides a landscape of NAFLD-HCC and its TME characteristics from a topological perspective. These findings indicate that targeting GPC3⁺/YAP⁺ cancer cells may improve the long-term prognosis of patients with NAFLD-HCC.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12935-026-04415-y.

Introduction

Hepatocellular carcinoma (HCC) is the third most common cause of cancer-related death globally [1]. The risk factors for HCC include chronic hepatitis B and C, alcohol addiction, metabolic liver disease (especially nonalcoholic fatty liver disease, NAFLD), and exposure to dietary toxins such as aflatoxin and aristolochic acid [2]. With the increasing prevalence of obesity and metabolic syndrome, NAFLD is becoming among the most common chronic liver diseases worldwide [3]. As a hepatic manifestation of metabolic disease, NAFLD begins with the accumulation of fat in the liver and progresses to liver inflammation known as nonalcoholic steatohepatitis (NASH), liver fibrosis/cirrhosis and, ultimately, NAFLD-associated HCC (NAFLD-HCC) [4]. While viral hepatitis remains the leading cause of HCC cases in China, vaccination and antiviral treatment programs are changing the profile of HCC cases. Moreover, NASH is becoming an increasingly important contributing factor to HCC [5]. By 2030, NAFLD is expected to become one of the primary causes of HCC [3]. Compared with viral hepatitis, NAFLD-HCC is often diagnosed at an advanced stage and has a lower survival rate [6]. However, the mechanism regulating the immune microenvironment in NAFLD-HCC remains poorly understood.

In this study, we used imaging mass cytometry (IMC) to reveal the phenotypic and functional information of the tumor microenvironment (TME) in NAFLD-HCC. IMC is a new technology that employs metal tags for biomarker labeling and a laser ablation system coupled with a mass spectrometer for in situ high-resolution quantification of formalin-fixed and paraffin-embedded (FFPE) Sect [7]. Compared with multiplex immunohistochemistry (mIHC), IMC technology minimizes signal crosstalk, simplifies multichannel panel design, and eliminates interference from spontaneous tissue fluorescence. This enables the simultaneous detection of multiple protein markers in one scan at single-cell resolution and provides an ideal approach for observing the TME. The identification of the distribution, phenotype and interaction of immune cells through IMC can facilitate an understanding of the role of the immune system in the progression of NAFLD-HCC.

IMC panel markers were selected using the following process: First, we used the nuclear and cell membrane markers DNA and β-CATENIN for cell identification and cell segmentation, respectively; second, we employed cell characteristic markers for cell type annotation, such as CD3, CD4, CD8A, CD11C, CD16, CD20, CD68, PanKeratin, and GPC3; and finally, the remaining markers were used to characterize cell phenotypes. For example, we used markers such as KI67, HISTONE-H3, PD-L1, VISTA, E-CADHERIN, ADFP, and YAP to describe tumor phenotypes such as proliferation, immunosuppression, epithelial-mesenchymal transition (EMT), and lipid metabolism. All the IMC markers are listed in Table 1.

Table 1.

Basic patient information

Patient characteristics
Patients, n	56
Male, n	40
Female, n	16
Age, years (mean ± SD)	62.63 ± 11.78
Age range, years	21–81
Cancer stage (TNM)
IA, n	3
IB, n	26
IIA, n	23
IIIA, n	3
IIIB, n	1
Overall survival
Survival, months (mean ± SD)	73.83 ± 37.26
Survival range, months	0.3–168.3

The objective of this study was to elucidate the distribution characteristics, specific subgroup functions, and regulatory mechanisms of the immune microenvironment of NAFLD-HCC through IMC technology. Furthermore, we aimed to identify valuable phenotypes for the clinical prognosis of patients with HCC, thereby providing a new way to improve the efficacy of targeted therapy.

Results

IMC images accurately restored tissue structure and cellular spatial localization in liver sections

A schematic diagram of the IMC workflow is shown in Fig. 1A. We first scanned single metal-labeled continuous section samples from healthy donors to determine the staining specificity and sensitivity of each channel in the IMC. All the different marker staining results exhibited strong and clear signals, with no significant channel overlap (Fig. 1B).

Fig. 1.

Workflow of IMC and representative mass cytometry images from normal liver tissue. (A) A schematic diagram of the IMC workflow, including sample processing and data analysis. (B) Single-color staining of different markers on the same ROI of consecutive slices. The imaging results of various signal channels are presented in pseudocolor form. Scale bar: 100 μm. (C) (a) H&E staining image of normal liver tissue. Scale bar: 400 μm. (b) Mass cytometry images of normal liver tissue magnified 2 times. PanKeratin (yellow), αSMA (cyan) and COLLAGEN I (purple) were used to portray the structure of the liver tissue. *, interlobular vein; ^, interlobular artery; and $, interlobular bile duct. Scale bar: 200 μm. (c) T cells were labeled with CD3 (red), CD4 (blue) and CD8A (green). Scale bar: 200 μm. (d) Monocytes and macrophages were labeled with antibodies against CD14 (red), CD16 (blue), and CD68 (green). Scale bar: 200 μm. (e) NK, endothelial and B cells were labeled with antibodies against CD57 (red), CD31 (blue), and CD20 (green). Scale bar: 200 μm. (f) A landscape of the above cells. Scale bar: 200 μm

To clarify that our panel can effectively restore tissue structure and the immune microenvironment, we conducted an IMC pre-experiment on a control section of normal liver tissue. HE staining of serial sections revealed the presence of interlobular veins, interlobular arteries, interlobular bile ducts and central veins (Fig. 1C a). The use of markers such as pankeratin and alpha smooth muscle actin (αSMA) in IMC, which were specific for hepatocytes and liver mesenchymal cells, allowed for a more accurate delineation of the above tissue structure (Fig. 1C b). In conjunction with the distinctive markers of CD3, CD4, CD8a, CD14, CD16, CD20, CD57, and CD31, the number and spatial distribution of hepatocytes, endothelial cells, hepatic stellate cells (HSCs), and diverse immune cells could be unambiguously delineated (Fig. 1C c-f).

Single-cell phenotyping of IMC revealed specific insights into NAFLD-HCC

IMC scanning was conducted on tumor regions, border regions and adjacent regions of NAFLD-HCC tissues and liver tissues in the control group, as previously described. The results demonstrated the tissue characteristic structure and immune cell infiltration in their respective regions. For example, the tumor antigen GPC3 was highly expressed in tumor regions, fibrous tissue was enriched in border regions, a normal liver structure was observed in adjacent regions and control tissue, and adjacent regions exhibited more immune cell infiltration than tumor regions did (Fig. 2A).

Fig. 2.

Single-cell clustering and proportions of different clusters. (A) Main cell types in tumor, border, adjacent and normal tissues. Scale bar: 200 μm. (B) A total of 18 cell clusters determined by 34 z scores of mean marker expression are shown in heatmaps, with their frequency distribution patterns across different regions shown in a box plot on the left. (C) Proportion of different cell clusters in tumor regions, border regions and adjacent regions. T, tumor; B, border; A, adjacent. *: P < 0.05; **: P < 0.01; ***: P < 0.001; ****: P < 0.0001; ns: Not significant

A matrix of a total of 1,127,014 single-cell data points was obtained by IMC scanning and analysis of 157 tumor regions, 81 border regions and 129 adjacent regions of interest (ROIs). We defined different cell types on the basis of unsupervised clustering of feature marker expression and constructed a heatmap of 18 clusters with their proportions in different regions (Fig. 2B). These clusters included 9 clusters of cancer cells, 5 clusters of immune cells, 3 clusters of mesenchymal cells and 1 cluster of hepatocytes.

Although the algorithm was effective for cell segmentation (Supplementary Fig. 1A), the unavoidable cell layering of tissue sections (Supplementary Fig. 1B) and the coexpression of some markers in a variety of cells resulted in the grouping of cells with spatial proximity or close relationships into a cluster. For instance, cluster 4 represented a class of T and NK cell populations that interacted closely. If cluster 4 was further subdivided, then CD8⁺ T cells expressing CD3 and CD8a; CD4⁺ T cells expressing CD3, CD4, and CCR5; and NK cells expressing CD57 could be separated from each other (Supplementary Fig. 1C). Similarly, cluster 5 was characterized by enrichment of endothelial cells and B cells, with CD31 and CD20 serving as markers, respectively. This cluster also exhibited high expression of ADFP, VISTA, and VIMENTIN, which may potentially function in the pathological process of NAFLD-HCC [8–10]. Studies have shown that B cells were near follicular dendritic cells in tumors and in the vicinity of high endothelial venules. This structure is regarded as one of the tertiary lymphoid structures (TLS) within the TME [11]. Clusters 9, 10, 13 and 16 were cell groups that were predominantly composed of cancer cells with high GPC3 expression (GPC3^hi cancer cells), which exhibited distinct patterns of immune cell infiltration. In particular, cluster 10 contained dendritic cells (HLA-DR⁺, CD11C⁺) and CD8⁺ T cells (CD3⁺, CD8⁺), cluster 13 contained infiltrating macrophages (CD14⁺, CD68⁺), cluster 16 contained NK cells (CD57⁺, GRANZYME B⁺), and cluster 9 had relatively low immune cell infiltration. Notably, cluster 9 highly expressed not only GPC3 but also YAP. YAP is another biomarker commonly linked to hepatocarcinogenesis [12], highlighting its potential as a marker of malignancy. Cluster 2 included a collection of resident macrophages (Kupffer cells, CD16⁺, and CD68⁺) and tumor-associated macrophages (CD14⁺, CD68⁺, and CD163⁺). Clusters 3 and 18 were identified as myofibroblasts and HSCs, respectively, on the basis of their distinct expression levels of αSMA and COLLAGEN I. Cluster 17 included hepatocytes defined by high levels of PanKeratin and E-CADHERIN, which also contained NK cells (CD57⁺ and GRANZYME B⁺). Clusters 1, 11 and 15 were concentrated mainly in tumor regions and were regarded as clusters of cancer cells with low GPC3 expression (GPC3^lo cancer cells). All the GPC3^lo cancer cell clusters had low immune cell infiltration. Among the clusters, cluster 7 exhibited the lowest KI67 expression, indicating that it may represent a group of resting or apoptotic cancer cells with the slowest proliferation. Cluster 12 represented antigen-presenting cells because of their high expression of HLA-DR and VISTA. Clusters 6, 8, and 14 represented individual monocytes, naive CD8⁺ T cells, and NK cells, respectively.

Statistical analysis revealed significant variations in the proportions of the aforementioned clusters across tumor, border and adjacent regions (Fig. 2C, Supplementary Fig. 1D). Furthermore, within the comparable region, there were significant differences in the proportion of the same cluster observed in different patients (Supplementary Fig. 1E). These results suggest that the TME of NAFLD-HCC exhibits considerable heterogeneity not only across different regions but also among different individuals.

The TME within tumor regions, border regions and adjacent regions demonstrated significant heterogeneity

To demonstrate the differences in the TME among tumor, border and adjacent regions, a tissue sample encompassing all three regions was selected (Fig. 3). The use of DNA dye facilitated the exact localization of each cell. This revealed a dense distribution of cells in the tumor region, a less dense distribution in the adjacent region, and a sparse distribution in the border region (Fig. 3A). The main structure of the border region was characterized by abnormal deposition and colocalization of the fibrosis markers αSMA and COLLAGEN I (Fig. 3BC). As illustrated in the orange, green and blue squares in the middle part of Fig. 3G, with the border region serving as the boundary, the left side is designated the tumor region, while the right side is the adjacent region. As anticipated, cancer cells in the tumor region presented increased levels of the tumor antigen GPC3, indicating enhanced tumor malignancy (Fig. 3D). Compared with cancer cells, dedifferentiated hepatocytes in the tumor region presented elevated levels of epithelial cell markers, including PanKeratin and E-CADHERIN (Fig. 3EF). Morphologically, cancer cells and differentiated hepatocytes in the tumor region exhibited irregular clumping, while the adjacent region retained a discernible structure, including vessels, hepatic sinusoids and hepatic cords. In addition, many vacuolated steatotic hepatocytes exhibited macrovesicular steatosis, a hallmark pathological finding in patients with NAFLD (Fig. 3G). The signal intensities of the abovementioned markers in the three regions are shown at the bottom of Fig. 3G. A review of the statistical data regarding the expression levels of these biomarkers in the three regions across all patients consistently mirrored the trajectory depicted in Fig. 3 (Supplementary Fig. 2).

Fig. 3.

Mass cytometry images of a representative rectangular ROI that includes tumor regions, border regions and adjacent regions simultaneously on one screen. (A-F) Some key markers are presented in different colors on the same screen. Scale bar: 200 μm. (G) Merged image of the single-staining images mentioned above. Scale bar: 200 μm. The tumor regions, border regions and adjacent regions are indicated by rectangles of different colors. The expression intensity of the above markers at different positions on the screen is presented in the form of heatmaps

The spatial locations of resident macrophages (CD16⁺, CD68⁺), infiltrating macrophages (CD14⁺, CD68⁺), dendritic cells (CD11C⁺), NK cells (CD57⁺), B cells (CD20⁺), CD4⁺ T cells (CD3⁺, CD4⁺) and CD8⁺ T cells (CD3⁺, CD8A⁺) are indicated by the corresponding markers (Supplementary Fig. 3A-I). Furthermore, the expression of T cell differentiation markers (GRANZYME-B and FOXP3) and macrophage polarization markers (CD86 and CD163) reflected the immune status of the aforementioned cells (Supplementary Fig. 3J-M). A majority of immune cells (except for dendritic cells) were observed more frequently in the adjacent region but not in the tumor region (Supplementary Fig. 3A-I). Notably, the expression levels of several typical immune checkpoints and immunosuppressive markers, including FOXP3, PD-1, PD-L1, and LAG3, were also higher in the adjacent region than in the tumor region (Supplementary Fig. 3K, N-P). These findings indicate that immunosuppression occurs mainly in the peritumoral tissue rather than inside the tumor tissue. The distribution characteristic is representative because the same conclusion can also be drawn on the basis of the statistical results of the signal intensity of the above markers or the cell percentage of the corresponding cell clusters in different regions of NAFLD-HCC (Fig. 2C, Supplementary Fig. 2).

Single-cell analysis revealed the landscape of the NAFLD-HCC TME across different regions

To conduct a more comprehensive statistical analysis of the common distribution characteristics of the NAFLD-HCC TME, we randomly selected 1,000 cells from each cluster and conducted a t-distributed stochastic neighbor embedding (tSNE) analysis on the basis of the expression of multiple markers. The points on the tSNE map are labeled according to their original cluster (Fig. 4A). The localization of tumor cells, border cells and adjacent cells on the tSNE map exhibited notable discrepancies, indicating substantial differences in their characteristics (Fig. 4B). A comparison of the expression of different feature markers on the tSNE map revealed basic consistency with the trend shown in Fig. 3 in a continuous field of view (Supplementary Fig. 4). GPC3 was highly expressed in clusters 9, 10, 13, and 16, and these clusters were concentrated in the tumor region (Supplementary Fig. 4, plot 20). αSMA and COLLAGEN I were highly expressed in cluster 3, namely, myofibroblasts, which were concentrated in the border region (Supplementary Fig. 4, plots 13 and 35). PanKeratin and E-CADHERIN were highly expressed in cluster 17, i.e., in dedifferentiated hepatocytes, and the cluster was concentrated in the adjacent region (Supplementary Fig. 4, plots 10 and 14). Linear discriminant analysis (LDA) was performed on the proportion of each cluster of ROIs from the tumor, adjacent and border regions. The 95% confidence ellipses of the tumor region and the adjacent region showed a separate trend, indicating differences in the distribution of typical cell clusters between the two regions. The border region lay at the transitional position between them, which aligned with the cellular phenotypic characteristics of the tissue transition zone. Myofibroblasts were found to be negatively correlated with the proportion of cells in the T and NK cluster and the monocyte cluster (Fig. 4D). These findings suggest that specific stromal cells and immune cells may regulate each other. The T and NK clusters were positively correlated with the KI67⁻ cell cluster and negatively correlated with the GPC3^hi cancer cell cluster in terms of cell proportion (Fig. 4E). These findings suggested that T and NK clusters may be involved in inhibiting and killing GPC3^hi cancer cells.

Fig. 4.

Differences in cellular composition and function among tumor regions, border regions and adjacent regions. (A) Single-cell data were reduced with a tSNE plot that combined three different regions. Cells were named according to the aforementioned clustered annotations. (B) tSNE plots of tumor regions, border regions and adjacent regions. (C) LDA of tumor regions, border regions and adjacent regions based on the percentage of each cluster in different ROIs. The arrow indicates malignant changes in the sample tissue. (D-E) Statistical correlation analysis of cell proportions. Pa: Adjusted P value

Cellular neighborhood analysis helps to elucidate the relationship between TME characteristics and clinical prognosis

To further analyze the interactions between cells, we adopted the concept of the cellular neighborhood (CN) as the smallest functional cell group, as others have previously described [13, 14]. The pattern diagram of CNs is shown in Fig. 5A. Using an unsupervised clustering method to analyze all possible cell combinations within the CNs, we identified 9 distinct clusters (Fig. 5B). The data indicated that across all patients, CN7 and CN8 were the predominant CNs in the tumor and border regions, respectively. Moreover, CN1 and CN3 were abundant in adjacent regions (Fig. 5C, Supplementary Fig. 5). To visualize CNs as cellular functional units within the TME in a topological form, we labeled them with different colors on 3 representative ROIs of different regions in the form of Voronoi diagrams. A comparison of continuous slices labeled with hematoxylin-eosin (H&E) staining revealed that the CNs precisely matched the topological morphology of the special structures in each region. Additionally, the primary cellular composition aligned with our earlier descriptions of the aforementioned regions (Fig. 5D).

Fig. 5.

Correlations of CN with the prognosis of NAFLD-HCC. (A) Schematic diagram to illustrate the concept of the CN. Scale bar: 50 μm. (B) Nine different types of CNs were clustered according to the composition of the cells. (C) The proportion of each CN in the ROIs from the tumor regions, border regions and adjacent regions. Patient IDs are shown in grayscale, and sex is presented in binary colors. (D) Voronoi plots of representative ROIs in each region to demonstrate the spatial composition of CNs. H&E-stained images of continuous slices were used as controls. Scale bar: 200 μm. (E) Comparison of YAP and GPC3 expression between the high-CN7 and low-CN7 groups

We further explored the independent prognostic value of CNs and other risk factors (including TNM stage, age, and liver cirrhosis grade, as detailed in Fig. 6) in patients with NAFLD-HCC using univariate and multivariate Cox proportional hazards regression analyses. The results confirmed that poorly differentiated HCC (Edmondson-Steiner grading system) and a high proportion of CN7 were independent risk factors for poor prognosis in patients with HCC across all tissue regions (Fig. 6; Table 2). CN7 was enriched in cluster 9 (Fig. 5B), a GPC3^hi cancer cell cluster characterized by significantly high expression of YAP in addition to GPC3 (Fig. 2B). We sorted cases on the basis of the average CN proportion of the ROIs and divided them into high-CN and low-CN groups, with a median proportion as the threshold. The level of YAP was significantly greater in the high-CN7 group than in the low-CN7 group (Fig. 5E, left). Notably, CN7 was also enriched in GPC3^lo cancer cells. There was no significant difference in overall GPC3 levels between the CN7 high group and the CN7 low group (Fig. 5E, right). This may be because the proportion of GPC3^lo cancer cells was much greater in the CN7-low group than in the CN7-high group (Supplementary Fig. 1G).

Fig. 6.

Forest plot of risk factors associated with NAFLD-related HCC across all regions

Table 2.

IMC panel of all the markers

Immune	Function	Stroma
CD3	CCR4	β-CATENIN
CD4	CCR5	DNA1
CD8A	CCR7	DNA2
CD11C	GZMB	PanKeratin
CD14	HSITONE H3	E-CADHERIN
CD16	KI67	GPC3
CD20	FOXP3	COLLAGEN I
CD31	PD-1	YAP
CD45RO	PD-L1	VIMENTIN
CD57	HLA-DR	VEGFR
CD68	LAG3	ADFP
CD86		αSMA
CD163		VISTA

In addition, we performed stratified analyses to further clarify how relevant risk factors affect prognosis across different tissue regions. The results revealed that the pathological grade of HCC was consistently an independent risk factor that significantly influenced prognosis in all three regions (Supplementary Figs. 6–8, Tables 3, 4, 5 and 6). Furthermore, in border regions, the proportion of CN8 significantly differed in univariate Cox regression. However, its independent prognostic significance was lost after adjusting for potential confounding factors in the multivariate model. These findings suggest that this factor may affect prognosis through synergistic interactions with other clinicopathological factors (Supplementary Fig. 7, Table 4). In tumor regions, the proportion of CN7 was an independent risk factor (Supplementary Fig. 8, Table 5), which was consistent with the findings in all regions (Fig. 6; Table 2).

Table 3.

Univariate and multivariate Cox regression results for all regions

Variables	Univariate			Multivariate
	HR (95% CI)	P		HR (95% CI)	P
Size	0.999 (0.974, 1.025)	0.952
Differentiation	4.342 (1.714, 11.000)	0.002		6.994 (2.221, 22.035)	0.001
Venous invasion	1.120 (0.496, 2.529)	0.785
Fatty liver	0.864 (0.154, 4.858)	0.868
Hepatitis grading	1.837 (0.842, 4.012)	0.127
PIVKA-II	1.000 (1.000, 1.000)	0.882
TNM stage	1.785 (0.970, 3.283)	0.063		1.356 (0.611, 3.011)	0.454
BCLC stage	3.587 (1.001, 12.850)	0.0497		2.429 (0.547, 10.784)	0.243
Cirrhosis	3.287 (0.481, 22.450)	0.225
AFP (ug/L)	1.000 (0.999, 1.002)	0.476
CA199 (U/ml)	1.002 (0.990, 1.014)	0.792
CEA (ug/L)	0.802 (0.600, 1.072)	0.136
CN1%	0.203 (0.017, 2.417)	0.207
CN2%	1.038 (0.061, 17.520)	0.979
CN3%	0.176 (0.009, 3.480)	0.254
CN4%	0.097 (< 0.001,19.380)	0.388
CN5%	286.1 (< 0.001, > 999.999)	0.676
CN6%	0.004 (< 0.001, 59.360)	0.263
CN7%	5.063 (1.398,18.330)	0.014		5.525 (1.379, 22.132)	0.016
CN8%	3.085 (0.563, 16.910)	0.194
CN9%	1.202 (0.042, 34.340)	0.914

Table 4.

Univariate and multivariate Cox regression results for adjacent regions

Variables	Univariate			Multivariate
	HR (95% CI)	P		HR (95% CI)	P
Size	0.986 (0.961, 1.013)	0.305
Differentiation	3.712 (1.311, 10.510)	0.014		10.755 (2.183, 52.984)	0.004
Venous invasion	1.052 (0.412, 2.689)	0.916
Fatty liver	0.687 (0.109, 4.323)	0.689
Hepatitis grading	2.252 (1.035, 4.902)	0.041		2.006 (0.455, 8.839)	0.358
PIVKA-II	1.000 (1.000, 1.000)	0.365
TNM stage	2.206 (1.183, 4.112)	0.013		1.137 (0.570, 2.269)	0.715
BCLC stage	5.855 (1.572, 21.810)	0.008		3.850 (0.254, 58.485)	0.332
Cirrhosis	5.611 (0.992, 31.740)	0.051		4.954 (0.784, 31.289)	0.089
AFP (ug/L)	1.001 (0.999, 1.002)	0.474
CA199 (U/ml)	1.001 (0.989, 1.014)	0.878
CEA (ug/L)	0.820 (0.584, 1.150)	0.249
CN1%	0.081 (0.004, 1.633)	0.101
CN2%	0.719 (0.031, 16.470)	0.837
CN3%	0.354 (0.016, 7.787)	0.510
CN4%	1.654 (< 0.001, > 999.999)	0.930
CN5%	NA	NA
CN6%	NA	NA
CN7%	11.490 (0.617, 213.900)	0.102
CN8%	2.217 (0.451, 42.210)	0.597
CN9%	4.142 (0.096, 179.300)	0.460

Table 5.

Univariate and multivariate Cox regression results for border regions

Variables	Univariate			Multivariate
	HR (95% CI)	P		HR (95% CI)	P
Size	1.010 (0.981, 1.039)	0.511
Differentiation	3.491 (1.154, 10.560)	0.027		3.138 (1.113, 8.846)	0.031
Venous invasion	1.345 (0.602, 3.004)	0.470
Fatty liver	0.908 (0.179, 4.592)	0.907
Hepatitis grading	1.556 (0.667, 3.630)	0.307
PIVKA-II	1.000 (1.000, 1.000)	0.318
TNM stage	1.237 (0.687, 2.226)	0.479
BCLC stage	1.589 (0.538, 4.694)	0.403
Cirrhosis	NA	NA
AFP (ug/L)	1.000 (0.999, 1.002)	0.796
CA199 (U/ml)	0.999 (0.987, 1.010)	0.799
CEA (ug/L)	0.807 (0.616, 1.056)	0.118
CN1%	0.006 (< 0.001, 130.900)	0.320
CN2%	< 0.001 (< 0.001, > 999.999)	0.168
CN3%	0.021 (< 0.001, 2.542)	0.115
CN4%	< 0.001 (< 0.001, > 999.999)	0.307
CN5%	< 0.001 (< 0.001, > 999.999)	0.732
CN6%	< 0.001 (< 0.001, > 999.999)	0.245
CN7%	3.382 (0.202, 56.730)	0.397
CN8%	29.150 (1.601, 530.900)	0.023		24.878 (0.947,653.805)	0.054
CN9%	2.878 (0.019, 428.800)	0.679

Table 6.

Univariate and multivariate Cox regression results for tumor regions

Variables	Univariate			Multivariate
	HR (95% CI)	P		HR (95% CI)	P
Size	1.003 (0.977, 1.030)	0.823
Differentiation	6.564 (2.886, 14.930)	< 0.001		13.787 (4.413, 43.077)	< 0.001
Venous invasion	1.087 (0.481, 2.457)	0.840
Fatty liver	0.989 (0.151, 6.478)	0.990
Hepatitis grading	1.768 (0.807, 3.871)	0.154
PIVKA-II	1.000 (1.000, 1.000)	0.675
TNM stage	1.795 (0.986, 3.269)	0.056		1.336 (0.516, 3.457)	0.550
BCLC stage	3.680 (1.067,1 2.700)	0.039		2.602 (0.671, 10.086)	0.167
Cirrhosis	3.150(0.473,20.970)	0.236
AFP (ug/L)	1.000(0.999,1.002)	0.445
CA199 (U/ml)	1.025 (0.978, 1.075)	0.294
CEA (ug/L)	0.789 (0.587, 1.061)	0.117
CN1%	0.523 (0.015, 18.090)	0.720
CN2%	1.475 (0.118, 18.380)	0.762
CN3%	0.181 (0.004, 9.305)	0.395
CN4%	0.100 (0.001, 7.379)	0.294
CN5%	> 999.999 (< 0.001, > 999.999)	0.596
CN6%	0.006 (< 0.001, 4.139)	0.124
CN7%	8.558 (1.056, 69.350)	0.044		12.718 (1.087, 148.769)	0.043
CN8%	2.309 (0.066, 80.980)	0.645
CN9%	0.026 (< 0.001, 11.770)	0.242

Discussion

Although the molecular mechanisms underlying the pathogenesis of NAFLD-HCC remain incompletely elucidated [15], chronic inflammation and the TME are widely recognized as crucial factors [16]. In patients with HCC, complex interactions occur among immune cells, tumor cells, and stromal cells. These interactions result in heterogeneous TME characteristics [17].

Currently, immune checkpoint inhibitors (ICIs) represent a promising class of immunotherapies for treating various cancers, including HCC [18]. However, studies indicate that ICIs may be less effective for NAFLD-HCC than for viral hepatitis-associated HCC [5, 19, 20]. The heterogeneity of the TME plays a significant role in influencing the efficacy of ICIs in NAFLD-HCC [21].Traditional clinical signatures cannot effectively predict HCC recurrence due to tumor heterogeneity and consequently do not contribute to the improvement of long-term survival in patients undergoing radical hepatectomy for HCC.[22] While single-cell transcriptome sequencing can delineate different subsets of cells in HCC [22], it does not provide information on the spatial relationships among these cells.

We employed IMC scanning with 37 channels to analyze 129 ROIs from a cohort of 56 patients across the tumor, border, and adjacent regions. This approach provides a novel perspective for observing the TME of NAFLD-HCC and identifying potential clinical targets. We identified and categorized 1,127,014 individual cells into 18 major cell clusters. Notably, significant variations were observed in both the distribution of major cell clusters and the microstructures among the three regions.

In summary, our findings revealed that immune cells, including those with immunosuppressive phenotypes, were more prevalent in adjacent regions than in tumor regions of NAFLD-HCC tissues. These findings suggest that the border region plays a critical role in limiting the infiltration of various immune cells into tumor tissues in NAFLD-HCC.

CN is the smallest unit of cellular interaction, and we introduce this concept here to further explore the position-function relationship between cells. We defined nine major CNs on the basis of their predominant cell compositions, which included four CNs primarily composed of immune cells, three CNs dominated by tumor cells, one CN characterized by stromal cells, and one CN with minimal cell interactions.

To validate CN distribution as an independent prognostic factor, we performed multivariate Cox regression analyses adjusting for the established clinical variables. We found that the pathological differentiation grade of HCC was consistently a key independent risk factor affecting the prognosis of NAFLD-HCC across various tissue regions. Moreover, some traditional prognostic criteria, such as TNM stage and BCLC stage, are regarded as the “gold standards” for postoperative liver cancer. However, these factors failed to emerge as independent risk factors in our statistical analysis. This may be because most of the patients in our cohort had TNM stage I/II disease (only 5 patients had stage III disease, and no patients had stage IV disease) or BCLC stage 0/A disease (only 7 patients had stage B disease, one patient had stage C, and no patients had stage D disease), resulting in severe homogeneity in classification. Traditional liver cancer staging criteria typically serve as the foundation for identifying surgical indications in patients with HCC. However, using them to further evaluate the clinical outcomes of postoperative patients may be too coarse to effectively distinguish and predict prognosis. According to our analysis, other major clinical indicators, including tumor size, hepatitis grade, fatty liver grade, liver cirrhosis grade, and serum levels of AFP, CEA, CA-199, and PIVKA-II, cannot serve as independent risk factors affecting the prognosis of patients with NAFLD-HCC. In contrast, the novel biomarker CN7 stands out as a potentially effective prognostic indicator. CN7 consists mainly of a cluster of GPC3^hi cancer cells, cluster 9, with high YAP expression. We confirmed that YAP expression was higher in the CN7-high patient group than in the CN7-low patient group (Fig. 5E, left), suggesting that YAP may be a driving factor of CN7 formation.

YAP, as the core transcriptional coactivator of the Hippo signaling pathway, plays a central role in hepatocarcinogenesis by regulating cell proliferation, apoptosis, stem cell properties, invasion and metastasis, drug resistance, and metabolic reprogramming [12, 23]. Clinical studies have confirmed that YAP nuclear localization is significantly associated with poor prognosis in liver cancer patients and that its activity is regulated by multiple factors, such as hepatitis B virus infection and metabolic abnormalities [24, 25]. Liu et al. demonstrated in animal models that YAP expression is elevated during the progression of NASH, which is a transitional stage in the development of NAFLD-HCC, thus suggesting that YAP may be involved in the pathogenesis of NAFLD-HCC [26]. Currently, direct clinical evidence concerning whether YAP affects the prognosis of NAFLD-HCC is lacking, and the findings of this study are expected to fill this research gap.

We also compared the expression of another malignant tumor driver marker, GPC3 [27], in patients with high CN7 expression as well as in patients with low CN7 expression. Although the proportion of GPC3^hi cancer cells was greater in CN7-high patients than in CN7-low patients, there was no significant difference in overall GPC3 expression between the two groups (Fig. 5E, right). This may be attributed to the fact that CN7-low patients had a higher proportion of GPC3^lo cancer cells (Supplementary Fig. 1G), which compensated for the overall GPC3 level through quantity. In conclusion, we believe that YAP, but not GPC3, is more likely to be the primary driver molecule responsible for CN7 formation. In future work, we will design appropriate animal models to validate this hypothesis.

In addition, when univariate Cox models were used, we found that a high proportion of CN8 was an unfavorable prognostic factor in the border region. Although this effect was no longer significant in the multivariate Cox regression analysis (p = 0.054), the proportion of CN8 in the adjacent region may serve as a synergistic reference indicator for prognostic assessment. CN8 is enriched in myofibroblasts, which are activated HSCs and contribute to fibrogenesis in NASH [28, 29]. Although NAFLD-related HCC can occur even in the absence of cirrhosis in many cases [2, 30], liver fibrosis tends to be a long-term outcome in patients with NAFLD [31, 32]. Therefore, in combination with our findings, reversing fibrosis as a therapeutic target may benefit patients with NAFLD-HCC.

This study had several limitations. First, the sample size of clinical cases included was relatively small, which may have led to insufficient statistical power in some subgroup analyses, limiting the stability and generalizability of the results. Future studies with larger sample sizes and multicenter designs are warranted to validate our findings. Second, additional spatial transcriptomics, along with cellular and animal functional intervention studies, are needed to clarify the underlying mechanisms involved. These findings provide stronger experimental evidence to support our clinical observations.

In conclusion, we introduced the unique phenotype of the NAFLD-HCC TME and revealed changes between different tumor-related regions across 56 clinical cases. A type of CN characterized by GPC3⁺/YAP⁺ cancer cell enrichment was identified as a functional cellular unit associated with adverse prognosis in patients with NAFLD-HCC.

Materials and methods

Samples and clinical data of patients

We investigated the phenotypic heterogeneity of NAFLD-HCC with samples from a patient cohort (n = 56) with an average follow-up of 7.2 years. Patients were included/excluded on the basis of the following criteria for a definitive diagnosis of NAFLD-HCC: (1) The inclusion criteria were as follows: Patients with a histopathological diagnosis of hepatocellular carcinoma confirmed by histological morphology and immunohistochemical examination and patients definitively diagnosed with metabolic fatty liver disease (MASH) according to the NAS pathological scoring criteria (NAS > 4 points), as shown by the typical H&E staining image of pathological sections in Supplementary Fig. 1F. (2) The exclusion criteria were as follows: Daily alcohol consumption ≥ 80 g alcohol/day (men) or ≥ 40 g alcohol/day (women) for more than 5 years; positive for HBV serological markers; positive HCV antibody test; positive for autoimmune antibodies; positive for schistosomiasis etiological or immunological tests; and a past history of liver injury due to drug and toxin poisoning, including aflatoxin, aristolochic acid, and other similar agents; and chronic liver diseases due to other causes. We adopted a prospective study protocol, and the cohort design at the beginning of the study was based on the definition of NAFLD; therefore, this manuscript did not use the nomenclature for MASLD advocated by the American Association for the Study of Liver Diseases (AASLD) or the European Association for the Study of the Liver (EASL) in 2023 [33]. All patients underwent radical hepatectomy, and 17 of them experienced relapses after an average of 1.7 years (Table 1). Surgical resection adhered to the anatomical resection standard of complete removal of the entire hepatic segment/lobus containing the tumor. Samples with a volume of 15 mm×15 mm×15 mm were collected from each site. Samples obtained from nonnecrotic intratumoral areas were defined as tumor regions. Samples containing the tumor edge were designated as the border regions. Two cm outside the tumor edge were collected as adjacent tissues. Patients’ clinical data were obtained from pathology and laboratory tests at the time of their visit, and survival data were obtained from long-term telephone follow-up after surgery, with the time of death or the time of the last follow-up as the end point (Supplementary Tables 1, 2). All surgical specimens used in this research were obtained with approval from the Ethics Committee of the Third Affiliated Hospital of Naval Medical University (Approval ID: EHBHKY2024-K011-P001). Written informed consent was collected from donors before specimen collection. All experimental procedures followed the Declaration of Helsinki.

IMC section preparation

The tissue blocks were further cut to a thickness of approximately 2 mm and then fixed in 10% buffered neutral formalin solution (100 mL formaldehyde + 900 mL water + 4 g NaH₂PO₄ + 6.5 g Na₂HPO₄) for 48 h. The fixed tissue block was rinsed with running water for 2–10 h to remove formaldehyde. The tissues were soaked in 30%, 50%, 70%, 80%, 95% and 100% ethanol for dehydration and placed in each concentration of ethanol for 45 min to 1 h. The tissue block was then treated for transparency: the tissue block was immersed in equal volumes of a mixture of ethanol, xylene and pure xylene for 1 h each time, and the solution was changed twice when immersed in pure xylene. The transparent tissue blocks were soaked in equal amounts of a mixture of paraffin and xylene and pure paraffin. The temperature of the incubator was set at 60 ℃ and the pure paraffin was soaked three times. The melting points were 52 ℃, 54 ℃ and 56 ℃ and the wax was permeabilized for 15 min each time. The melted wax was poured into the embedding frame, the tissue block was clamped with prewarmed forceps and placed flat on the bottom of the frame with the section facing down, and the embedding frame was placed in cold water to rapidly cool and solidify and remove after 30 min. A microtome (Leica RM2245) was used to cut the wax block into 4 μm thin slices, the slices were spread in 45 ℃ water, and a glass slide was used to pick up the slices, which were dried slightly at room temperature and then placed on a 65 ℃ toaster for 1 h, after which they were heated in a 60 °C constant temperature oven for 2 h.

FFPE sections were dewaxed by soaking in xylene at 60 ℃ for 20 min, after which one change of fluid occurred. The sections were then soaked in 100%, 95%, 80% and 70% ethanol at room temperature for 5 min each. The sections were incubated in 1x antigen retrieval solution (Agilent, s236784-2) at 96 ℃ for 30 min. After the samples were allowed to cool to room temperature (70 ℃), ddH₂O and phosphate-buffered saline (PBS) were used for rinsing for 10 min each. Blocking was performed with 3% bovine serum albumin (BSA)/Dulbecco’s phosphate-buffered saline (DPBS) for 45 min at RT. The antibody mixture was formulated with DPBs containing 0.5% BSA. After the blocking solution was aspirated from the sections, the antibody mixture was incubated overnight at 4℃. The next day, the sections were washed twice with DPBS containing 0.2% Triton X for 8 min each time and then washed twice with DPB for 8 min each time. To correct for signal overflow between adjacent channels, the slides were stained with single metal-tagged antibodies in accordance with the principles of compensation. In the antibody mixture, some antibodies were metal direct labeling antibodies (Fluidigm), and the remaining antibodies needed to be further incubated with secondary antibodies after primary antibody incubation to bind the metal tag. The method is summarized as follows: The secondary antibody of the metal tag was incubated for 1 h at room temperature (recommended dilution: 1:50 ~ 1:100), the samples were washed twice with DPBS containing 0.2% Triton X for 8 min each time, and the samples were then washed twice with DPBs for 8 min each time. The sections were counterstained with Iridium DNA Intercalator solution (201192 A; Fluidigm, 1:400) in DPBS for 30 min at RT. Afterward, the sections were rinsed with ddH₂O for 5 min and air-dried for at least 20 min before IMC scanning. The specificity and optimal working concentration of each antibody were determined by single staining of experimental and control samples. See Supplementary Table 3 for the complete metal tag antibody panel.

IMC scanning

Experienced pathologists identified 2–5 ROIs for IMC scanning under white field microscopy, with each ROI confirmed to exhibit typical pathological features of its corresponding region. The procedure was performed using a Hyperion imaging system (Fluidigm). Each ROI was scanned at a resolution of 1000 × 1000 μm and files were saved in MCD format for subsequent image analysis.

H&E staining

FFPE tissue samples were placed in hematoxylin staining solution for 5 min, rinsed with running water for 5 min, differentiated with 1% hydrochloric acid-ethanol for 30 s, and rinsed with running water for 30 s, followed by 0.2% ammonia aqueous solution back to blue for 1 min. The samples were subsequently rinsed with running water for 30 s, stained with 0.5% eosin staining solution for 1 min, rinsed with running water, heated in a constant temperature oven at 40 ℃ for 1 h, dried and sealed with medium resin.

IMC analysis

Tiff files of each ROI were generated from the IMC raw data using the MCD Viewer (v1.0.560.6). Nuclear masks were generated using Ilastik (v1.4.0) pixel classification based on iridium nuclear staining (Fluidigm, 201192 A). Cell masks were generated by CellProfiler (v4.2.6) using β-catenin as a cell membrane indicator. Maps of cell masks and signaling channels were fed into histoCAT (v1.76) to generate a matrix for each ROI at the single-cell level, containing information on cell location, neighboring cells and marker expression [34]. The expression data were censored using a 95th percentile cutoff to exclude outliers. Single-cell data were ultimately analyzed using R (v4.3.3). PhenoGraph clustering and heatmap visualization were performed using the R packages Rphenograph (v1.5.0) and Pheatmap (v1.0.12), respectively. Statistical t-distributed stochastic neighbor embedding (t-SNE) [35] was used for dimensionality reduction visualization through the R package Rtsne (v0.16) in R language. LDA was performed on the proportion of each cluster of ROIs using the MASS (v7.3) package in the R language as previously described [36].

Cell neighborhood analysis

Cellular neighborhood analysis was employed to explore functional units within a single-cell interaction network [13, 14]. In accordance with the established definition, the cell neighborhood was defined as all cells within a specified radius window around each central cell. A fixed-diameter window positioned centrally within a specific cell was deemed to encompass all the cell sets included within it and thus to constitute a single CN. Given that the median diameter of all the cells counted in this study was 8 μm, the diameter of the window was set at 16 μm. This ensured a close interaction between the central cell and its surrounding cells within an 8 μm distance, categorizing them into the same CN. Neighboring cell information was automatically generated by the histoCAT pipeline. In our images, one pixel equals 1 μm in length. Cells within an 8-pixel radius (8 μm) of the central cell were considered to belong to a CN. A heatmap was generated by undifferentiated clustering of R for assessing which CN each cell belongs to. We then create Voronoi plots on the basis of this information and count the proportion of CNs within each ROI.

Statistical analysis

The raw data were analyzed as described above. The processed data obtained from histoCAT were then entered into GraphPad Prism software (10.1.2) for statistical analysis and graph construction. The normality of the data was determined using the Shapiro-Wilk test. For pairwise comparisons, Student’s t test was performed for normally distributed data. For non-normally distributed data, the Mann-Whitney U test was applied for independent samples, and the Wilcoxon signed-rank test was applied for paired samples. For multiple-group comparisons, repeated-measures ANOVA was applied for normally distributed data, followed by Tukey’s HSD test for multiplicity adjustment in post hoc pairwise comparisons. For nonnormally distributed data, the Friedman test followed by Dunn’s post hoc test with Bonferroni correction for P value adjustment was applied accordingly. The error bars of all the quantitative data in the bar graph represent the median ± 95% confidence interval (median ± 95% CI). Spearman’s rank correlation analysis was performed using the cell cluster proportion matrix across all patients to evaluate pairwise correlations between distinct cell clusters. The false discovery rate (FDR) method was applied for multiple testing correction. An adjusted P value < 0.05 was considered to indicate statistical significance, and a correlation coefficient of |r|>0.4 was defined as the threshold for a biologically relevant correlation between two clusters.

To explore the impact of CN and other potential risk factors on the clinical prognosis of patients with HCC, univariate and multivariate Cox proportional hazards regression analyses were performed. The primary endpoint of prognosis is overall survival (OS). All potential prognostic factors (including CN proportion) were first included in the univariate Cox regression analysis to screen out variables with statistical significance (P < 0.1). The variables with significant differences identified by univariate analysis were subsequently incorporated into the multivariate Cox regression model to adjust for confounding effects among various factors, and the hazard ratio (HR) and 95% CI were calculated to clarify the independent prognostic value of CN and other risk factors for patients with HCC. In addition, considering the potential bias caused by different numbers of ROIs within each region, a mixed effect model was adopted for data analysis to eliminate the interference of ROI quantity on the study results, ensuring the reliability and robustness of the statistical conclusions. Stratified analysis was conducted according to tumor, border, and adjacent regions to further clarify the prognostic value of relevant factors in different tissue regions. The raw data for this part of the statistical analysis are shown in Supplementary Table 4.

Author contributions

Conceptualization, X.W., G.Z., Y.G. and W.W.; project administration, W.W.; supervision, W.W.; methodology, X.W.; investigation, X.W.; validation, X.W., J.H.; resources, X.W. X.S.; data curation, X.W., B.X. and S.C.; formal analysis, B.X., Q.J. and X.Y.; software, Q.J.; visualization, B.X., Q.J.; writing-original draft, B.X., W.B. and X.G.; writing-review and editing, B.X., W.W.; funding acquisition, X.G., G.Z., Y.G. and W.W. All the authors have read and agreed to the submitted version of the manuscript.

Funding

W.W. was funded by the National Natural Science Foundation of China (82472914).Y.G. was funded by the National Natural Science Foundation of China (82574152) and the Sanming Project of Medicine in Shenzhen (SZSM202311019).X.G. was funded by the National Natural Science Foundation of China (82203323).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Competing interests

The authors declare no competing interests.