Single cell profiling of endothelial cells (ECs) from normal liver
To gain baseline information on liver ECs, we performed single cell transcriptome profiling on naïve normal liver from immune-competent C57BL/6 and immune-deficient SCID (C.B-17 scid) mice. Our single cell data were derived from multiple 10× Genomics sequencing runs. We benchmarked and observed that batch effect was minimal and rarely impacted cell clustering result (see method section for more info on minimization and monitoring of batch effect). A total of 1446 cells from normal C57BL/6 mouse livers were mapped into ten clusters representative of various leukocyte cell types (cluster 1–6), epithelial cells (cluster 7) and ECs (cluster 0, 8–9) (Fig. 1a and Supplementary Fig. 1a–c). Further, re-clustering of 471 ECs identified five subpopulations within the EC population (Fig. 1b and Supplemental Table 1). Consistent with the structure of the liver vasculature, we were able to annotate these subpopulations with consensus markers as central vein (CV) ECs (cluster 2, using Rspo3 as a marker) [7], sinusoidal ECs (SEC) (cluster 0, 1 and 3, using Clec4g as a marker) [8, 9] and portal vein (PV) ECs (cluster 4) (Fig. 1c). Using gene expression patterns across subpopulations together with a few well-established EC zonation markers such as Rspo3 and Bmp2, we were able to spatially assign the three SEC clusters along the PV to CV axis, representing EC zonation (Fig. 1c). Analysis of single cells collected from normal livers of naïve SCID mice aligned with our findings from C57BL/6 mice regarding EC subpopulations and subpopulation-specific marker genes (Supplementary Fig. 1d–e).
To assess whether there are mouse strain-specific differences in naïve liver ECs, we performed cell clustering on combined ECs from C57BL/6 and SCID livers (Fig. 1d). ECs from C57BL/6 and SCID mice aggregated into the same CV or PV clusters (cluster 3 and 4, respectively) (Fig. 1d and Supplementary Fig. 2a). However, within each CV or PV cluster, ECs segregated by strain. Sinusoidal ECs were again classified into three clusters, which also sub-segregated by strain: one cluster (cluster 2) was SCID-derived, whereas the other two clusters (cluster 0, 1) were from C57BL/6 mice. Strain-specific genes were found in liver ECs (Supplemental Table 1 and Supplementary Fig. 2b) with more differentially detected genes in SEC than PV or CV, due to higher SEC cell counts. Functional enrichment analyses of these differential genes showed that genes highly expressed in SECs of C57BL/6 mice were significantly associated with immune-related pathways, such as antigen presentation and processing, complement cascade and hematopoietic cell lineage (Fig. 1e). For example, expression of genes involved in MHC class II antigen presentation including Cd74, H2-Ab1, Ctsb, and Lgmn were significantly higher in SECs derived from immune-competent mice (C57BL/6) compared to SCIDs. On the other hand, genes more highly expressed in the sinusoidal ECs of SCID mice showed enrichment in ribosomal genes and oxidative phosphorylation. These findings suggested that the immune status may contribute to the transcriptional profile of liver ECs.
To benchmark the EC subpopulation and zonation genes derived from this study, we compared the results to recently published human [10] and mouse [8] studies. While the conservation of zonation genes was limited between mouse and human liver ECs, a number of reported mouse liver EC zonation profiles [8] did exhibit similar zonation patterns in our dataset, such as Sox17 and Dll4 as periportal EC markers (Supplementary Fig. 2c). However, the expression patterns of many reported zonation genes [8] were not reproduced in our current study. To further validate our zonation findings, we performed RNAScope for validation of marker expression in the respective specialized liver vessel structures. As expected Rspo3 was highly expressed in the CV [7] (Supplementary Fig. 2d). Additionally, we were able to confirm Selp as another CV-enriched gene and Nrg1 as a PV-specific gene (Fig. 1f).
Intrahepatic tumor ECs formed a distinct subpopulation, and adjacent normal ECs were affected by the presence of tumor
To induce liver cancer in situ in immune-competent mice, we employed a hydrodynamic delivery (HDD) approach targeting oncogenic pathways in hepatocytes. Activated Kras (mKrasG12D) and deletion of p53 (CRISPR-sgTrp53) were delivered to immune-competent C57BL/6 and BALB/c mice as well as immune-compromised SCID mice, via HDD of plasmid DNA. Mice subjected to HDD developed multiple tumors throughout the liver. Histological examination of these tumors revealed a mixed hepatocellular/cholangiocarcinoma phenotype (Fig. 2a), similar to what has been reported in some liver cancer patients [11]. Certain tumor regions showed a solid trabecular structure, as typically observed in human hepatocellular carcinoma, whereas other regions displayed bile duct differentiation features, as evidenced by the positive staining for cytokeratin 19, and featured a Masson-positive stromal reaction, similar to human cholangiocarcinoma. Tumor-bearing livers were either used undissected (resulting in a mix of cells from tumor and adjacent normal tissue) or dissected macroscopically into tumor and adjacent normal tissue. It is noteworthy that the dissected adjacent normal portion of a liver might have contained small tumors that were not visible by gross inspection. Similarly, dissected tumor tissue might have contained small portions of adjacent normal tissue.
When comparing ECs from HDD-induced liver tumors to those from normal liver, we initially focused on one strain of immune-competent mice. A total of 356 single ECs from tumors or adjacent normal tissues of C57BL/6 mice were identified and combined with the previously collected 471 ECs from naïve normal liver for cell-clustering analyses (Fig. 2b). As previously, ECs from CV and PV formed distinct clusters (cluster 3 and 4), but cells segregated by sample type within the CV cluster. Sinusoidal ECs were separated into two clusters: one cluster (cluster 0) predominantly composed of cells from naïve normal liver; the other (cluster 1) mainly composed of cells from tumor-adjacent normal liver. ECs from liver tumors formed a distinct cluster (cluster 2). Examination of cluster-specific marker genes showed that tumor ECs expressed a unique set of genes such as Col18a1 and Aplnr, but also shared a number of expressed genes with venous ECs, especially with PV, such as Cd63, Ehd4, and Cd200 (Fig. 2c and d, Supplementary Fig. 3a, and Supplementary Table 2). In contrast, tumor EC transcriptomes showed minimal overlap with sinusoidal EC transcriptomes. As previously reported, significantly higher numbers of genes were detected in tumor ECs than normal ECs (Fig. 2e) [12]. Genes preferentially expressed by tumor ECs were enriched in functions including angiogenesis, cell proliferation, cell-cell adhesion, and response to cytokines (Fig. 2f).
To determine whether the presence of an adaptive immune system affected the pattern of tumor EC gene expression, HDD-induced liver tumors from immune-deficient SCID mice were also profiled. As seen in immune-competent mice, a distinct cluster of tumor ECs was observed in HDD-induced tumors in SCID mice, along with similar tumor EC-specific genes and their associated functions (Supplementary Fig. 3b–d).
Many gene expression differences were observed between naïve liver and so-called adjacent normal liver tissue, suggesting that the tumor influenced the transcriptome of adjacent normal stromal cells including ECs. For example, Selp was increased in SECs and induced in PV from tumor-bearing livers. This result was supported by RNAScope findings (Figs. 1f and 2d). Another molecule Lrg1, a mitogen demonstrated to promote angiogenesis in the presence of TGF-β1 [13], was highly expressed in tumor ECs and upregulated in adjacent normal ECs (Fig. 2g). The presence of tumor also induced upregulation of Sema3d, which encodes the ligand for plexin D1 and is involved in angiogenesis, in adjacent normal ECs (Fig. 2g). We also observed that the transcriptome alterations in tumor-adjacent normal vs. naïve normal liver differed between C57BL/6 and SCID mice (Supplementary Table 2), although there was overlap in certain differential genes. For example, upregulation of Serpina3h was only observed in the liver SECs of C57BL/6 mice (Fig. 2c). With the same cutoffs applied, a lower number of differential genes between adjacent normal and naïve normal liver were detected in CV ECs compared to SECs (Supplementary Table 2). However, there were fewer CV and PV cells than SECs, which limited the power of this type of analysis. In summary, the presence of a liver tumor led to transcriptome changes in ECs of the adjacent normal liver region.
Minimal impact of host immune status on tumor EC transcriptome
Since normal ECs from C57BL/6 and SCID mice showed substantial transcriptional differences, we combined ECs derived from both tumor-bearing and normal (naïve and adjacent normal) livers from C57BL/6, BALB/c and SCID mice and re-performed cell clustering (Fig. 3). Over 1600 ECs clustered primarily based on EC subtypes, such as cluster 2 and 4 representing CV and PV ECs, respectively. SECs roughly formed three clusters representing sinusoids from naïve normal C57BL/6, naïve normal SCID, and adjacent normal liver tissues from all three mouse strains (Fig. 3a–c). Again, tumor ECs collectively formed a distinct cluster (Fig. 3b). Within the tumor EC cluster, cells from different strain backgrounds mixed well and did not display disparity (Fig. 3c). Nevertheless, we queried for differentially expressed genes in tumor ECs between immune-competent (C57BL/6 or BALB/c) and immune-compromised (SCID) strains and found only a few genes with significant scores (adjusted P value < 0.01 and fold change > 2) (Fig. 3d and Supplementary Table 2). Additionally, fewer genes overlapped between the two comparison results. Thus, it is reasonable to conclude that intrahepatic tumor ECs were minimally impacted by immune status of the host tissue.
Minimal impact of intrahepatic tumor type on tumor EC transcriptome
To determine whether the tumor type affects tumor EC phenotype, we utilized a second intrahepatic tumor model, where human HT-29 colorectal cancer cells, which we used previously for s.c. tumor studies [1], were implanted directly into the liver parenchyma. Such a model has been used to represent metastatic colorectal cancer [14]. Intrahepatic HT-29 tumors grew as a single mass in the liver and displayed features of a well differentiated CRC, as evidenced by neoplastic cells with a glandular differentiation and mucinous material in the glandular lumens, as well as a prominent stromal compartment (Supplementary Fig. 4a). To increase the power of the analysis, ECs from all sample collections including naïve normal liver, adjacent normal liver, and intrahepatic HDD-derived and HT-29 tumors were merged and subsequently clustered into nine subpopulations (Fig. 4a). At first glance, two clusters stood out because they were almost exclusively composed of cells from tumors (Fig. 4b). Therefore, these clusters were designated as tumor ECs. While CV- and PV-derived ECs showed one distinct cluster each, the SEC subpopulation was further divided into three clusters despite a shared common gene signature: one cluster was almost exclusively comprised of cells from naïve livers (only collected from C57BL/6 and SCID mice), which was named sinusoid.naïve, one cluster mainly contained ECs from tumor-bearing livers (adjacent normal or mix) and thus named sinusoid.adjacent, and one cluster designated as sinusoid.intermix contained cells from both naïve and tumor-bearing livers (Fig. 4a). Additionally, two new distinct clusters emerged. They were annotated as arterial ECs and lymphatics based on known marker genes (Fig. 4a left). For example, Stmn2 and Sox17 were expressed in arterial cells [1] and Mmrn1 and Pdpn are lymphatic-specific [1]. We further validated lymphatic-specific expression of Tbx1 by RNAScope (Supplementary Fig. 4b). Lymphatic ECs, which were mostly tumor-associated (Fig. 4b), showed a higher gene count per cell (Fig. 4c) than ECs from normal liver.
The cumulative 506 tumor ECs derived from intrahepatic tumors (HDD-induced and HT-29 tumors) separated into two distinct subpopulations (Fig. 4a). The top subpopulation-specific genes in clusters tumor.EC1 and tumor.EC2 highly overlapped with those previously identified as “tip-like” and “stalk-like” EC marker genes in s.c. tumors, respectively (Fig. 4d) [1]. For example, Dll4 and Notch4 expression was limited to tip-like cells, Tgfbr3 was expressed in stalk-like cells, while both EC populations expressed similar levels of Kdr (Vegfr2) (Supplementary Fig. 4c).
Within the two tumor EC clusters, cells from various mouse strains and tumor types (mouse and human tumors) were well-intermixed in the t-SNE plot (Fig. 4b). Focusing on only SCID mice, we further checked gene expression differences in tumor ECs between HDD-derived and transplanted HT-29 intrahepatic tumors. Only a few genes showed preferential expression in one or the other intrahepatic tumor model (Supplemental Fig. 4d). Thus, the tumor type exerted minimal influence on the tumor EC transcriptome profiles in liver.
Integrated analysis of normal and tumor ECs from different organs
The clear separation of tumor ECs from naïve/adjacent normal ECs, and the observed conservation of tip- and stalk-like gene signatures in tumor ECs from s.c. and liver tumors, prompted us to look for commonalities that could serve as tumor EC-specific markers. Consequently, we performed an integrated analysis by combining our liver and heart [1] data with published single cell data on ECs from normal mouse tissues including lung [15] and kidney [16] obtained using comparable single-cell sequencing techniques. Even when sequence read coverages in ECs from different sources were comparable, the number of genes detected could vary significantly (Supplementary Fig. 5a). ECs from different normal organs (heart, lung, kidney, liver) clustered based on tissue of origin (Fig. 5a). Most normal organs further displayed EC subpopulations, such as liver (cluster 0, 2, 4, 11), kidney (cluster 7, 8) and lung (cluster 5, 9), whereas heart ECs formed one cluster (cluster 1) at the applied clustering resolution. At the applied cluster resolution, lymphatic ECs from heart and liver mapped into one cluster (cluster 10), indicating substantial similarity, which was further supported by a high-correlation coefficient score in transcriptome (Supplemental Fig. 5b). Tumor ECs clustered away from normal ECs. Despite conservation of tip- and stalk-like genes, s.c. tumor ECs formed a separate cluster (cluster 3) from intrahepatic tumor ECs (cluster 6).
Because the ECs in the combined analysis above were pooled from different sources, batch effects should not be ignored. However, because ECs in the s.c. tumor cluster (cluster 3) were collected from different xenograft models and were profiled in separate 10× Genomics single cell runs, the most influential factor on EC transcriptome appears to be the host tissue. Each cluster displayed distinct tissue-specific and subpopulation-specific gene signatures (Fig. 5b), which were deprived of the usual batch-related genes such as Jund, Klf and ribosomal genes. Although common tumor EC genes such as Nid2 and Col15a1, were detected in s.c. and liver tumor-derived ECs, two distinct clusters formed based on the site of tumor growth (i.e., s.c. and liver). Collectively, ECs express common endothelial lineage markers but also carry tissue- and cell state-specific imprints (proliferation, metabolism, etc.).
Following up on this observation, we focused on HT-29 tumors in SCID mice and assessed differentially expressed genes in ECs derived from intrahepatically and subcutaneously grown tumors. Many genes were preferentially expressed in intrahepatic tumor ECs, whereas few genes were uniquely expressed in s.c. tumor ECs (Fig. 5b). Although the majority of the liver-specific genes appeared to be involved in house-keeping functions (Fig. 5c and Supplementary Table 3), there were a number of genes such as Tgfb1, Hras, Sumo2 and Sox17 that are involved in cellular functions beyond housekeeping (Fig. 5d). These results showed that the host organ, but not the tumor type, exerted significant influence on EC phenotypes in liver tumors. Taken together, tumor ECs carry conserved as well as host organ-imposed gene signatures.
Chimeric myeloid-endothelial cell type present in tumor-bearing livers
When assessing EC marker gene expression in all single cells collected from the liver, we noticed that a subpopulation of the Kupffer cell cluster also expressed conventional EC marker genes (Supplemental Fig. 6a–c). Based on gene and UMI counts, these cells did not appear to be doublets (Supplemental Fig. 6d). In support of the existence of chimeric cells, double-positive cells expressing the EC marker CD31 (Pecam1) and the Kupffer cell marker Clec4f were detected by IHC (Fig. 6a). These cells were often detected close to blood vessels without integrating into the vessel. To further investigate these cells, we combined all Kupffer cells and ECs for re-clustering. Two subpopulations (cluster 7 and 8) emerged between ECs (cluster 0, 2, 3, 4 and 5) and Kupffer cells (cluster 1 and 6), which displayed mixed molecular phenotypes (Fig. 6b). In particular, cells in cluster 7 expressed a higher number of genes, which is often seen in tumor ECs, and were more prevalent in tumor-bearing livers than naïve livers (P value < 0.001 by Fisher Exact test) (Fig. 6c). The host mouse strain did not affect the presence of these two populations (Fig. 6c). All cells from clusters 7 and 8 expressed EC marker genes, such as Pecam1, Clec4g, Rspo3, Egfl7, and Robo4, as well as monocyte/macrophage markers, such as Csf1r and C1qa (Fig. 6d and Supplementary Fig. 7a).
Furthermore, there were differences between cluster 7 and 8. In particular, cells in cluster 7 showed preferential expression of monocyte/macrophage markers F13a1 and Itgam (Cd11b), whereas cells in cluster 8 expressed higher levels of Kupffer cell markers Clec4f and Vsig4 (Fig. 6d and Supplementary Fig. 7b, c). Further trajectory and cluster tree analyses suggested that clusters 7 and 8 possessed more EC transcriptome properties than Kupffer cell properties (Supplementary Fig. 7d). Altogether, this led us to name cluster 7 “macrophage-EC” and cluster 8 “Kupffer-EC”. To further investigate the myeloid signal in these chimeric ECs, we first classified all myeloid cell populations and subpopulations in our dataset (Supplementary Fig. 8 and Supplementary Table 4). Next, we constructed a trajectory including all myeloid cells (except plasmacytoid dendritic cells, i.e., pDC) and the chimeric ECs, which revealed that the “Kupffer-ECs” (red) were most closely related to naïve and inflamed Kupffer cells (teal and turquoise), whereas the “macrophage-ECs” (pink) clustered with macrophages (different shades of blue), all of which were located at the same end of the trajectory. A closer look at the EC features of “Kupffer-ECs” and “macrophage-ECs” revealed that specific CV and PV marker genes described above were barely expressed in either subpopulation (Fig. 6d; Rspo3 as an example of a CV marker). This observation suggests that chimeric ECs represent an intermediate phenotype between Kupffer cells or macrophages and sinusoidal ECs.