2.1 Dynamic landscapes of the cellular composition and characterization during DEN-induced hepatocellular cancer

To gain insights into the immune cell landscape during HCC tumorigenesis, we performed scRNA-seq to profile the transcriptomes of liver cells from DEN-induced HCC mouse models. We used a total of four tissue samples for this purpose. Two of these samples were whole liver tissues obtained from two individual mice that were 6 months old. The other two samples were obtained from one mouse that was 16 months old, where one sample was a liver tumor and the other was adjacent nontumorous liver tissue. It has been well known that visible liver tumors may be observed in 7−8 months after DEN intraperitoneal injection on 2-week-old mice.^{25, 26} Thus, in this study, we collected livers from 6 and 16-month mice after DEN injection, respectively, where the livers of 6-month-old mice have no visible tumor but existing inflammatory cells by microscopy, belonging to a precancerous stage, whereas ones from 16-month-old mice developed overt tumors with liquefactive necrosis, being in late-stage (Figure S1). After the quality control, 17,493 genes and 8821 cells were used for downstream analysis (Table S1). Then, we are able to assign these cells into 10 main cell groups, including T cells, neutrophil cells, B cells, hepatocytes, endothelial cells, epithelial cells, DCs, fibroblasts, and erythrocytes, based on these gene markers (Figures 2A and S2), showing that the TME is highly complex with different amount and status of immune cells. In this HCC model, in addition to hepatocytes and erythrocytes (0.2%) that might residual in the livers, there was a large amount of immune cell (78.9%) (Figure 3A) infiltration during HCC tumorigenesis and development.

We then performed the unsupervised clustering on each cell types using the spectral clustering method implemented in Seurat.²⁹ Through our refined analysis, we were able to identify a total of 36 subgroups (Figure 2B) with T cells having the largest number of eight subgroups (Figures 3B and F and Table S2), illustrating its diversity characteristics. Here, except for B cells with three subtypes (Figures 3C and G and Table S2), DCs (Figures 3D and H) and neutrophils (Figures 3E and I) were also divided into seven and six subsets, respectively, indicating various responses under different stimulations. Interestingly, the composition of main cell types was similar between tumor and its adjacent (Figure 3A), but the subgroups belonging to DCs and neutrophils were distinctly different (Figures 3H and I and Table S3). T cells were the major immune cell cluster in the HCC TME, with the sample mean of 36.4% (Figure 3A). The mean frequencies of B cells (25.9%), neutrophil (8.6%), hepatocyte (7.9%), DC (6.7%), endothelial (10.1%), epithelial (1.8%), monocyte (1.3%), fibroblast (1.1%), and erythrocyte (0.2%), respectively (Figure 3A), suggesting that, besides the original resident, a large number of immune cells infiltrated in this pathological environment.

Furthermore, we compared the cell type compositions between early and late stages during tumorigenesis (Table S4). In particular, the proportions of T and B cells significantly increased in the late stage (p value < 0.001; see Materials and Methods) (Figure 3A), whereas the percentage of neutrophils, hepatocytes, endothelial cells, epithelial cells, and fibroblasts decreased during the process, showing cell composition was changing from early to late stages. In the late stage, the total number of T and B cells had exceeded the half of the total number of cells (Figure 3A), where certain subtypes belonging to T and B cells, DCs, and neutrophils were obviously increased in late stages (Figures 3F–I). The collective data, as expected, showed that the immune cell composition and characteristics change during HCC tumorigenesis (between the early and late time points), as indicated by Figure 3, displaying the landscape of the multicellular clusters with different distribution patterns in the HCC microenvironment.

2.2 The presence of malignant hepatocytes in the early stage before overt HCC occurrence

To investigate the malignant transformation status of hepatocytes during HCC tumorigenesis, we first assigned all hepatocytes from the above livers into four subgroups according to their gene expression patterns (Figure 4A), and then performed the pseudotime trajectory analysis on these cells (Figure 4B), showing that the four subgroups were respectively located at different development or differentiation stages. We further compared the hepatocytes across early and late stages (Figure 4C), showing that the majority of hepatocytes from early and late stages were classified into distinct subgroups. As expected, a group of hepatocytes characterized by Afp-positive cell subsets is accumulated in late stage, especially in tumors (Figure 4D). Interestingly, however, a few hepatocytes from early stage were clustered in the hepatocyte-3 subgroup (Figure 4C), which gathered the hepatocytes from tumor of late stage. Moreover, some of these cells also expressed Afp, a well-known HCC marker (Figure 4D), suggesting that the malignant hepatocytes could exist in the early stage before overt HCC occurrence.

Furthermore, the pseudotime trajectory analysis on hepatocytes from early and late stages showed the continuous process (Figure 4E), suggesting that some of the hepatocytes from the early stage might evolve to tumor formation, along with the switch from Afp-negative status to Afp-expressing status during tumorigenesis (Figure 4F). In the switching process, albumin may keep the same level among different subclusters of hepatocyte (Figure 4G), but Apoa2 and Epcam could increase or decrease, respectively (Figures 4H and I). The expression of Afp was dramatically increased in subcluster hepatocyte 3 rather than in other three subclusters. The expression of Apoa2 was higher in subcluster 0 and 3 than in other subclusters. The expression of Epcam was highly expressed in subcluster 3, which is different from the expression patterns of Afp, Alb, and Apoa2. In previous studies based on bulk sequencing, Afp is often negative in the precancerous stage. In this study, we demonstrated that some cells in the early stage could also express Gpc3 at a single cell resolution, along with the liver stem cell marker Met highly expression (Figures 4J and K), where the specific mode of the expression of these stem/progenitor markers might be an indication of early cancer occurrence.

2.3 T and B cells exhibit the differential subpopulations between early and late stage during tumorigenesis

We analyzed the subpopulations and their functional activities of the immune cells, especially T cells with eight subgroups and B cells with three subsets (Figure 5). The results demonstrated the distinct diversity of immune features in precancerous and cancerous lesions. Among eight T-cell subgroups (Figure 5A, left), except for one reflecting CD4⁺ T cells, seven represented different CD8⁺ T cell subpopulations with known unique signature genes (Figure 5A and Table S2).^30-32 Differential composition analysis revealed that the proportion of T cell subtypes in early and late stages during HCC tumorigenesis was distinctly different (Figure 5). In particular, by fitting a fixed effect generalized linear model (GLM), we found that the proportions of exhausted T cell (T0), MAIT (T1), and Irf8+ T (T5) were significantly increased in the late stage during tumorigenesis (Figures 5C and E), whereas naïve T cell (T3), NKT (T4), and CD4⁺ T (T6) were decreased in late stage (Figures 3A and 5 and Table S3), implying that T cell functions could be suppressed during HCC formation. Three B-cells subgroups represented naïve B (B0), memory B cells, and plasmablasts, respectively (Figure 5F). There was a clear increase in the proportion of naïve B cell (B0) from the early to late stages during tumorigenesis; by contrast, the proportions of plasmablast (B1) and memory B cell (B2) were dramatically decreased in the late stage (Figure 5H and Table S4), confirmed by the GLM analysis with significant interaction effects between B cell subgroups and stages (p < 0.001) (Figures 5H and I). The dramatic change in B cell maturation process suggested TME may inhibit the maturation of naïve B cells during tumorigenesis.

We further validated these findings in a cohort of 365 HCC patients (Table S5). By determining the association between marker genes of naïve B cells (TRAV1-2, CD161), MAIT cells (SLC4A10, ZBTB16, RORC), and NKT cells (GZMA, GZMK, GZMB, CXCR6) and the survival rate, we discovered that higher expression levels of these genes were correlated with better survival rates. The detailed results can be found in Figure S3. These results doubly confirm our findings at the gene expression level, making our findings more reliable and convincing.

2.4 Diverse phenotypes of intratumors PD-1⁺ T cells

We observe that T cell populations within tumors are distinguished by different patterns of environmental characteristics. Our study revealed that tumor-resident T cells may be exposure to different levels of hypoxia, inflammation, and nutrient deprivation. Although many of these responses (e.g., hypoxia or activation) individually represent phenotypic continuums, their combinations may lead to more dissociated states. Figure 6 shows that regulatory T cells (T6) were mostly seen in the CD4⁺ subpopulation and were defined by the coexpression of CD25 and Foxp3. PD-1⁺ cells were predominantly observed within CD8⁺ subpopulations, particularly in exhausted CD8⁺ T cells (T0) and naive T cells (T3). The T0 subset, characterized by the highest level of PD-1 expression among CD8⁺ T cells, also exhibited positivity for other coinhibitory receptor Havcr2, activation marker CD38, and costimulatory receptors such as Tnfrsf9 and Icos. This phenotype might be correlated with anti-PD-1 treatment response and exhausted T cells.^{33, 34} Subpopulations with PD-1 levels similar were different in the expression on activation markers and costimulatory receptors. Cluster T3 had the same expression pattern as T0 but the activation and costimulatory markers were at a lower level. PD-1⁺ clusters (T2, T5, and T7) were characterized by a lack of Havcr2 and heterogeneity in the expression of markers such as Tnfrsf9, CD38, and Ctla-4. Unlike CD4⁺ Regulatory T cells (T6) presented in the early stage, PD-1⁺ clusters were mainly present in the late stage, but neither of them show a significant difference between tumor and adjacent tissue (Figure 6). All above demonstrated a diversity phenotypic among PD-1⁺ T cells present in tumorigenesis. Previous findings showed a correlation between the heterogeneity of T cells and the efficacy of anti-PD-1 therapies.^{35, 36} This provides a clue as to why some immunotherapies are ineffective in the treatment of HCC. We boldly speculate that these different types of PD-1⁺ T cells may interact with liver cells and other types of immune cells, sustaining an inflammatory environment and promoting disease progression.

2.5 Intercellular signaling interaction network could promote HCC development

In order to understand the interactions among different cell types within the HCC microenvironment n DEN-induced mice, we employed a ligand–receptor interacting repository by the diffusion of secreted molecules (www.CellPhone DB.org).^{37, 38} We investigated the expression levels of ligands and receptors in each cell type and determined the ligand–receptor pairs that displayed significant cell-type specificity (Figure S4). This could predict interactions between cell populations via specific ligand–receptor binding and generate a potential cellular communication network in the HCC microenvironment. B0-naïve B cells, hepatocytes, all subtypes of the endothelial and DC6 (Ccl17⁺ DC), which are dominant in the late stage, harbored the highest “outgoing” signals. Hepatocyte 3 (late stage Hepatocyte) and dysfunctional T cells harbored the highest “incoming” signals (Table S3). All subtypes of epithelial are harbored both “outgoing” and “incoming” signals. The results demonstrated that B0-naïve B cells could bind Ccr2 on the dysfunctional T, by releasing cytokine Ccl2, Ccl24, Ccl8, Ccl11, and bind Cxcr3 by releasing cytokine Ccl20. In this way, naïve B cells could regulate T cells. Meanwhile, hepatocyte3 (late stage hepatocyte) could bind Igf1r with B0-naïve B cells, by releasing cytokine Igf1. Briefly, the hepatocytes in late stage could regulate B0-naïve B cells. More interesting, naïve B cell binds Ctla4 with all types of T cells except T7-effect T cell, by releasing CD86. Figures 7 and S5 showed an outline of selected ligand–receptor interactions, indicating that T cells are likely the “terminal,” receiving single from other cells, especially from endothelial, epithelial, and hepatocytes. B cells are more like “originating,” they could release signal to other cells, at the same time, they can “self-regulation” (Table S3). Altogether, our data support potential regulation mechanisms of different cell types in HCC TME.

2.6 Cell interaction analysis reveals B cell activation in both mouse and human samples

We identified eight basic cell types, including T cells, monocytes, hepatocytes, B cells, CD34-pre-B cells, NK cells, BM cells (basement membrane cells), and platelets (only in blood) in three samples from three HCC patients (Figures 8A and B). The clinical information of the three patients is presented in Table S6.) We found that monocytes and NK cells accounted for a greater proportion of tumor cells compared with normal cells (Figure 8C). We tried to build an intercellular signaling interaction network by mapping to the ligand–receptor interacting repository from CellPhone DB (www.CellPhone DB.org).³⁷ This time we tried interaction between different cell types and showed the most significant ligand–receptors (Figure 8D). We found high interactions between immune cells, including monocytes and CD34-pre-B cells, via molecules such as ALB (ALBINO3-like protein 1), FcRn complex (IgG receptor FcRn large subunit p51), GRN, and TNFPSF1A.

2.7 The markers related to naïve B cells are correlate with better prognosis in HCC patients

Our finding revealed statistically significant variations in the proportions of multiple cell types between the two sample groups (Figure 9), such as memory B cell (B2), dendritic cell (DC2), and Afp+ hepatocyte (H3). Some clusters share patterns with the bulk sequencing data (Figure 9A) tested by Wilcoxon Rank Sum test. We observed a relevance between the proportion of cell types with survival for immune cells. In particular, we discovered that patients with a higher proportion of MAIT (T1), exhausted CD8⁺ T cell (T0), and naïve B cell (B0) have a better survival rate, while a lower proportion of CD38⁺ T cell (T2), naïve T cell (T3), and Irf8+ T cell (T5) associated with a higher survival rate (Figure 9B). All clues demonstrated the proportion of unmatured B cells could be a key regulator of inflammatory responses in the TME and a potential biomarker of the therapeutics for the HCC.

4.1 DEN-induced HCC mouse model

We utilized a DEN-induced HCC mouse model, employing C57BL/6 mice procured from the Shanghai Laboratory Animal Center. Our study included three DEN-induced mice, consisting of two 6-month-old mice and one 16-month-old mouse, all pathologically diagnosed with HCC. To induce HCC, we administered an intraperitoneal injection of 25 mg/kg DEN on the twelfth day after birth. Subsequently, we harvested whole liver tissues from the 6-month-old DEN-induced mice and paired fresh tumors along with adjacent liver tissues from the 16-month-old DEN-induced mouse.

4.2 Tissue dissociation and single cell suspension preparation

Fresh tissue samples were initially cut into approximately 1 mm³ pieces. Subsequently, they were gently homogenized by repeatedly pressing through a 70 µm cell strainer (BD) into a solution of phosphate-buffered saline (PBS) (Invitrogen) containing 10% fetal bovine serum. The residual material on the 70 µm cell strainer was collected and subjected to a 20-min digestion process with 0.05% collagenase IV on a shaker at 37°C. The resulting mixture was then filtered through a 40 µm cell strainer into the previously obtained cell suspension. Following this, the suspended cells were centrifuged at 200×g for 5 min. After discarding the supernatant, the pelleted cells were exposed to erythrocyte lysis buffer (Solarbio) and incubated on ice for 5 min to remove red blood cells. The cell pellets were subsequently resuspended in a buffer solution after being washed twice with 1 × PBS.

4.3 Hematoxylin and eosin staining

Slides were prepared by heating them on a slide warming table for 5−10 min, with the treatment being repeated if necessary. Subsequently, the tissue sections were hydrated through a graded alcohol series, with each concentration held in decreasing order for 1−2 min (starting from 100% alcohol and decreasing to 50%). The slides were then washed in tap water.

For staining, the sections were immersed in hematoxylin for 3−5 min and rinsed under running tap water for a brief period, not exceeding 5 min. Differentiation was achieved by placing the slides in 1% acid alcohol (1% HCl in 70% alcohol) for 5 min. (Microscopic examination confirmed the nuclei to be dark purple or reddish-blue, while the tissues appeared pale.) Following differentiation, slides were rinsed in running tap water until sections reverted to a blue hue, achieved by immersion in an alkaline solution followed by another tap water rinse.

Eosin staining was performed for 30 s, followed by a tap water rinse lasting 1−5 min. To complete the process, dehydration was carried out through a series of increasing alcohol concentrations, and clarity was achieved by immersion in xylene. The stained slides were then observed under a microscope.

4.4 Library preparation

Detailed procedure refer to the User Guide of 10× Genomics V2 libraries preparation available at: W020190620596403691451.pdf (ion.ac.cn).

Subsequently, the libraries were subjected to sequencing on an Illumina HiSeq 2000 platform (paired-end configuration; read 1: 26 cycles; i7 index: 8 cycles; i5 index: 0 cycles; read 2: 98 cycles). The Drop-Seq method was executed in accordance with the McCarroll Lab's protocol, and you can find comprehensive operational instructions in Drop-seq—McCarroll Lab.

4.5 Low level processing and filtering

For the critical processing steps, Cell Ranger software was employed, including demultiplexing, alignment, barcode counting, and unique molecular identifier (UMI) counting. To ensure data quality, cells with fewer than 100 UMIs were excluded. Subsequently, Seurat was applied for data normalization, dimensionality reduction, unsupervised clustering, and visualization of single-cell transcriptomes.

4.6 Quality control and preprocessing

We employed the k-means clustering technique to categorize cells based on their total expression and mitochondrial proportion. The integration of multisample data and batch effect correction was accomplished using the Harmony software. Moreover, to ensure the accuracy of downstream analyses, we employed the DoubletFinder software to identify and eliminate potential doublets. Following quality control, we utilized 17,493 genes and 8821 cells for downstream analysis.

4.7 Clustering and cell type identification

To delineate and characterize cell types in our dataset, we initiated a clustering analysis using Seurat with the first 20 principal components (PCs). Subsequently, we conducted differential expression analysis employing MAST and differential variability analysis using the Bartlett test to pinpoint marker genes specific to each cluster.

4.8 Cell type differential composition analysis

We used the scdney R package (https://github.com/SydneyBioX/scdney) to examine the relationship between cell type proportions and mouse phenotypes. Subsequently, we applied a fixed-effects GLM to the number of cells within each cell type.

4.9 Trajectory analysis

To explore the trajectory of hepatocyte cell development during tumorigenesis and its progression, we employed the Monocle2 software.

4.10 Cellular communication analysis

We conducted an analysis inspired by Vento-Tormo et al.³⁷, utilizing the ligand-receptor pairs available in the CellPhone DB database. To assess cell-cell interactions, we employed permutation tests on the product of the average gene expression levels of ligands in one cell type and receptors in another.

4.11 Integration with TCGA data and survival analysis

We leveraged single-cell RNA-seq data to deconvolute cell types within TCGA LIHC bulk RNA-seq datasets using dtangle. Subsequently, we applied a Cox proportional hazards model to evaluate the association between cell type proportions and survival outcomes.

4.12 Statistical analysis

Statistical analyses were conducted using R version 4.2.0. The first 20 PCs of the datasets were clustered using Seurat. Differential expression analysis was subsequently performed employing MAST, and differential variability analysis was conducted using Bartlett's test to identify marker genes specific to each cluster.