Single-Cell Meta-Analysis Uncovers the Pancreatic Endothelial Cell Transcriptomic Signature and Reveals a Key Role for NKX2-3 in PLVAP Expression

Abstract

BACKGROUND:

The pancreatic vasculature displays tissue-specific physiological and functional adaptations that support rapid insulin response by β-cells. However, the digestive enzymes have made it difficult to characterize pancreatic endothelial cells (ECs), resulting in the poor understanding of pancreatic EC specialization.

METHODS:

Available single-nuclei/single-cell RNA-sequencing data sets were mined to identify pancreatic EC-enriched signature genes and to develop an integrated atlas of human pancreatic ECs. We validated the findings using independent single-nuclei/single-cell RNA-sequencing data, bulk RNA-sequencing data of isolated ECs, spatial transcriptomics data, immunofluorescence, and RNAScope of selected markers. The NK2 homeobox 3 (NKX2-3) TF (transcription factor) was expressed in HUVECs via gene transfection, and the expression of pancreatic EC-enriched signature genes was assessed via RT-qPCR.

RESULTS:

We defined a pancreatic EC-enriched gene signature conserved across species and developmental stages that included genes involved in ECM (extracellular matrix) composition (COL15A1 and COL4A1), permeability and barrier function (PLVAP, EHD4, CAVIN3, HSPG2, ROBO4, HEG1, and CLEC14A), and key signaling pathways (S1P [sphingosine-1-phosphate], TGF-β [transforming growth factor-β], RHO/RAC GTPase [guanosine triphosphatase], PI3K/AKT [phosphoinositide 3-kinase/protein kinase B], and PDGF [platelet-derived growth factor]). The integrated atlas revealed the vascular hierarchy within the pancreas. We identified and validated a specialized islet capillary subpopulation characterized by genes involved in permeability (PLVAP and EHD4), immune-modulation (FABP5, HLA-C, and B2M), ECM composition (SPARC and SPARCL1), IGF (insulin-like growth factor) signaling (IGFBP7), and membrane transport (SLCO2A1, SLC2A3, and CD320). Importantly, we identified NKX2-3 as a key TF enriched in pancreatic ECs. DNA-binding motif analysis found NKX2-3 motifs in ≈40% of the signature genes. Induction of NKX2-3 in HUVECs promoted the expression of the islet capillary EC-enriched genes PLVAP and SPARCL1.

CONCLUSIONS:

We defined a validated transcriptomic signature of pancreatic ECs and uncovered their intratissue transcriptomic heterogeneity. We showed that NKX2-3 acts upstream of PLVAP and provided a single-cell online resource that can be further explored by the community: https://vasconcelos.shinyapps.io/pancreatic_endothelial/.

Highlights.

• A unique analysis framework was used to circumvent ambient RNA contamination inherent in pancreatic transcriptomic data to identify a biologically relevant pancreas endothelial cell (EC)–enriched signature for pancreatic ECs across species and development stage, which included genes involved in ECM (extracellular matrix) formation (COL15A1, COL4A1, and HSPG2), permeability (PLVAP, EHD4, CAVIN3, HEG1, and CLEC14A), and key signaling pathways: S1P (sphingosine-1-phosphate), IGF (insulin-like growth factor), TGF-β (transforming growth factor-β), RHO/RAC GTPase (guanosine triphosphatase), PI3K/AKT (phosphoinositide 3-kinase/protein kinase B), and PDGF (platelet-derived growth factor).

• An integrated atlas of human pancreatic ECs from 3 pancreatic data sets was introduced in this study that uncovered the pancreatic vascular hierarchy including an islet and exocrine capillary subpopulation, which was validated in adult human bulk RNA-sequencing data, fetal human spatial transcriptomic data, and in situ in embryonic day 18.5 mouse pancreas.

• The NK2 homeobox 3 (NKX2-3) TF (transcription factor) was a key pancreas EC-enriched signature gene of interest. Motif analysis demonstrated ≈40% of the pancreas EC-enriched signature genes had a DNA-binding motif for NKX2-3 in their transcription start sites. NKX2-3 induction demonstrated regulation of PLVAP, a marker of endothelial fenestration, and other signature genes.

• We have made the pancreatic EC-enriched differentially expressed genes and integrated human pancreatic EC atlas available for further exploration online (https://vasconcelos.shinyapps.io/pancreatic_endothelial/).

Endothelial cells (ECs) are the interface for tissue-blood exchange of oxygen, nutrients, and metabolic waste products. ECs also secrete soluble factors, known as angiocrine factors, that are important for tissue morphogenesis, repair, and homeostasis.¹ Furthermore, ECs adopt unique tissue-specific characteristics to meet the functional needs of the surrounding parenchyma.^2,3 The recent use of single-nuclei/single-cell RNA-sequencing (sn/scRNAseq) technologies has allowed major advances in the characterization of tissue-specific EC signatures for various organs.^4,5 However, characterization of pancreatic ECs is challenging as the exocrine pancreas secretes high levels of hydrolytic enzymes that prevent proper isolation of islet ECs. This also results in pancreatic single-cell RNA-sequencing (RNAseq) data sets to be prone to high levels of ambient RNA contamination due to cell lysis. Thus, there is lack of pancreatic ECs in published single-cell RNAseq atlases such as the murine endothelial atlas.⁶ In addition, the previously described organotypic signature for pancreatic ECs identified ambient acinar-specific genes as the top pancreatic EC signature genes, such as CPA1, CLPS, CEL, CELA3B, CTRBB1, and PNLIP.^7,8 Furthermore, most of pancreatic data sets have sequenced isolated, in vitro cultured islets,^9–17 which have low number of ECs. Therefore, these data sets do not allow for dissecting intratissue pancreatic EC heterogeneity (eg, exocrine versus endocrine) or allow for extracting information regarding islet EC specialization compared with other tissue ECs.

In this article, we have analyzed multiple sn/scRNAseq, bulk RNAseq, and spatial transcriptomics data sets to identify a specialized pancreatic EC-enriched gene signature and introduced an integrated human pancreatic EC atlas. This allowed us to profile the transcriptomics of specialized intraislet capillary ECs. Importantly, we identified the NK2 homeobox 3 (NKX2-3) TF (transcription factor) regulates pancreatic EC-enriched signature genes. Our approach provides a comprehensive overview of the intertissue and intratissue transcriptomic landscape of pancreatic ECs, which can help us better understand the role of islet microvessels in regulating glucose homeostasis and advance research in diabetes therapy. This article is also accompanied by a publicly accessible resource to interactively explore the integrated atlas of human pancreatic ECs and pancreatic EC-enriched signature genes (https://vasconcelos.shinyapps.io/pancreatic_endothelial/).

Materials and Methods

Data and Code Availability

All data are available in the main text or the Supplementary Material. The code to recreate figures and analysis pipelines can be found at https://github.com/vasconceloslab/pancet/. Please see the Major Resources Table in the Supplemental Material.

Data Acquisition

Raw count matrices and metadata for the sn/scRNAseq data sets used in this article were derived from their respective online databases/repositories. For the Descartes gene expression atlas (https://descartes.brotmanbaty.org/), sparse gene (row) by cell (column) matrices along with cell annotation were downloaded for all pancreatic cells and all ECs. For the Tabula Muris gene expression atlas, processed Seurat objects for all fluorescence-activated cell sorting sequenced organ data sets were downloaded, and the raw count matrices and metadata available for the cells were extracted (https://figshare.com/projects/Tabula_Muris_Transcriptomic_characterization_of_20_organs_and_tissues_from_Mus_musculus_at_single_cell_resolution/27733). Similarly, for the adult and neonatal pancreatic data from the study by Tosti et al, processed Seurat objects were downloaded (http://singlecell.charite.de/cellbrowser/pancreas/) and the raw count matrices and available metadata were extracted from the objects. For the Tabula Sapiens atlas, processed Seurat objects containing all tissue EC data and pancreatic cell data were downloaded (https://figshare.com/articles/dataset/Tabula_Sapiens_release_1_0/14267219), and then the raw count matrices and available metadata were pulled from the objects. Bulk RNAseq counts were obtained from (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE157546). Spatial transcriptomics Visium output files and deconvolution data were obtained from (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE197317).

Seurat Analysis

Data sets were analyzed using Seurat (v5.0.1).¹⁸ Low-quality cells with <200 unique genes and genes with 0 counts across all cells were removed. Cells with high unique feature counts (potential doublets or multiplets) above a certain threshold were omitted. The thresholds were determined for each data set individually based on distribution of feature counts. Cells with high mitochondrial content were filtered out. Data were log normalized, and highly variable genes were detected using variance-stabilizing transformation (vst). Cell cycle scores were generated for each cell using Seurat markers for S phase and G2M phase. Scoring followed the strategy outlined by Tirosh et al¹⁹ using the AddModuleScore function in Seurat. Principal component analysis (PCA) was conducted on scaled log-normalized data of the top 2000 most variable genes after cell cycle scores, and percentage of mitochondrial genes was regressed out. The top 30 principal components (or as determined using the elbow plot to assess contribution of PCs [principal components] to overall variance) were used for nonlinear dimension reduction using uniform manifold approximation and projection. The first 30 principal components were used to generate the Shared Nearest Neighbors graph, and clustering was conducted using the Louvain algorithm. Spatial transcriptomics data were analyzed using the SCTransform pipeline.²⁰

Deriving Signature Genes

Differential expression (DE) analysis was conducted between populations of interest with the FindMarkers function from Seurat using the Wilcoxon rank-sum test for EC atlases and pancreatic data sets and MAST for the integrated atlas. For the curated set of EC genes, significant differentially expressed genes (DEGs; log2 fold change [log2FC], >0.5; Bonferroni-adjusted P<0.05) were identified for ECs against other cell types excluding lymphatic ECs and stellate cells for all pancreatic data sets. For pancreatic EC-enriched genes from EC atlases, significant DEGs (log2FC, >0.25; Bonferroni-adjusted P<0.05) were identified for pancreatic ECs against all other available tissue ECs. For vascular subpopulation signatures from the integrated atlas, significant DEGs (log2FC, >0.25; Bonferroni-adjusted P<0.05) were identified for all subpopulations. This was then filtered using any genes that were significantly differentially expressed in pancreatic ECs (log2FC, >0.25; Bonferroni-adjusted P<0.05) across at least 2 EC atlases. Subpopulation signature genes were empirically optimized for specificity to their respective subpopulation within the integrated atlas itself and to pancreatic ECs in the human adult pancreatic data set from the study by Tosti et al (Tosti-Adult) and fetal pancreatic data set from Descartes. Initial signatures demonstrate low specificity of the exocrine capillary signature. Therefore, capillary signatures were further filtered with DEGs identified from DE analysis conducted specifically between capillary subpopulations (log2FC, >0.25; Bonferroni-adjusted P<0.05). The initial signature for precapillary arteriole demonstrated specificity to stellate cells; therefore, genes enriched in stellate cells COL4A1 and MCAM were removed from the signature. We then assessed Spearman correlation for all the vascular subpopulation signature scores with both islet and exocrine proportions for each tissue sample. We evaluated the relative difference in correlation of the signature scores between islet and exocrine regions using the Fisher z test of the Spearman coefficients. A higher Fisher z test score for a particular subpopulation signature indicates a stronger correlation with islet proportions relative to its correlation with exocrine proportions, thus serving as a quantitative measure of whether a signature is enriched within the islet or exocrine region.

Comparisons of DEGs Between Different Data Sets

For comparisons between genes using different versions of the same gene name, gene sets were updated to their latest approved symbols by comparing to the Hugo Gene Nomenclature Committee database as of April 2023. For comparison between mouse genes and human genes, all mouse genes were converted to their human homologs using the R homologene package (v1.4.68.19.3.27); for genes that did not have a human homologue, the case of the gene name was converted. To handle duplicates and multiple alternative symbols, similarity of symbols was compared based on Levenshtein distance using the stringdist (v0.9.12) package, and the most similar symbols were kept.

Integrating Data Sets

To generate the integrated EC atlas, the pancreatic EC population from the Descartes fetal pancreatic data set and the adult and neonatal pancreatic data set from the study by Tosti et al were merged. The EC subpopulation from the pancreatic data sets was individually processed using the Seurat pipeline detailed above. After clustering, contaminating clusters were identified by comparing DEGs between clusters. Clusters with differentially upregulated nonvascular markers such as endocrine and exocrine pancreatic cell markers were considered contaminating clusters. Once contaminating cells were removed, remaining ECs from the different data sets were merged into 1 Seurat object using the merge function. The merged object was processed using the Seurat pipeline as detailed above, except the source of the data set was regressed out while scaling. The data were integrated with Harmony (v1.0)²¹ to further remove variability from the source of the data set (Descartes database or from the study by Tosti et al) followed by the individual donor. The resulting integrated or harmonized principal components were used for uniform manifold approximation and projection dimension reduction and clustering as detailed above.

SCORPIUS Trajectory Analysis

Pseudotime trajectory analysis was performed using the R package SCORPIUS (v1.0.8).²² Trajectory inference was conducted on uniform manifold approximation and projection embeddings using the infer_trajectory function. For determining the top genes involved in the ordering of the cells along pseudotime, the top 50 markers were identified in each cluster in the integrated pancreatic EC atlas through DE analysis as described above. Only the log-normalized counts for these genes were retained to form a truncated expression matrix. The gene_importances function of SCORPIUS was then used to determine genes, whose expression level across the pseudotime best matches the time order using the gene_importances function.

Reference Mapping

To map a query data set to the reference data set, anchors were identified using the FindTransferAnchors function in Seurat, and then the annotations from the reference data set were projected onto the query using MapQuery.

Bulk RNAseq Analysis

Log normalization of counts per million was conducted on raw counts. Z scores were obtained by scaling the log normalization of counts per million values using the ScaleData function in Seurat. PCA was conducted using log normalization of counts per million values. DE analysis on exocrine and endocrine samples was conducted using DESeq2 using the raw counts. Module score was derived using log normalization of counts per million values using the AddModuleScore function in Seurat.

Gene Set Overrepresentation Analysis

The gene set overrepresentation analysis was performed using the package EnrichR (v.0.2.2)^23–25 using the following data sets: WikiPathway (2023), Reactome (2022), KEGG (Kyoto Encylopedia of Genes and Genomes; 2021), BioPlanet (2019), Panther (2016), and BioCarta (2016). Terms with adjusted P<0.05 or intersection size <l (ie, only l gene is present for that term) were removed. Afterward, the terms were manually examined to remove duplicates and keep terms of interest. The gene set enrichment analysis was visualized using iGraph (v1.6.0) and visNetwork (v2.1.2).

TF Motif Analysis

We look for the presence of NKX2-3 motifs in the promoter region of the list of signature genes using the HOMER software (v.4.1.1).²⁶ It scans for the given motif in the promoter region of all the transcripts in a given genome.

Data Visualization

Uniform manifold approximation and projections, dot plots, box plots, and volcano plots were generated using SCPubr (v1.1.2).²⁷ Hierarchical clustering was generated using ggtree²⁸ (v3.4.4) except for the heatmap for pseudotime trajectory analysis, which was generated using the draw_trajectory_heatmap function of SCORPIUS. Heatmap for bulk RNAseq analysis was generated using heatmaply (vl.4.2). Venn diagrams were plotted using ggvenn (v0.1.9). Scatterplots were generated using ggplot2 (v3.4.1). Breaks in violin plots were created using ggbreak (v0.1.1).²⁹ Normalized percentage difference between 2 populations p1 and p2 was calculated by dividing the percentage difference by the percentage expression in p1 and multiplying by 100.

Immunofluorescence Staining

Tissue cross sections were obtained from E18.5 mouse pancreatic tissue. Pregnant CD1 dams were euthanized via CO₂ asphyxiation, in accordance with IACUC (Institutional Animal Care and Use Committee) procedures at UT Southwestern Medical Center. Embryos were dissected in ice-cold PBS, and pancreata were obtained for fresh-frozen embedding. Pancreata were briefly dabbed on a kimwipe and washed 1× in optimal cutting temperature media. Pancreata were then placed in optimal cutting temperature–filled embedding and held over liquid nitrogen until frozen. Samples were then stored at −80 °C. Samples were cryosectioned by UT Southwestern Medical Center histology core. Briefly, slides were placed in 4% paraformaldehyde at room temperature for an hour and then washed 3× in PBS for 5 minutes per wash. Slides were permeabilized with 0.3% Triton X-100 for 12 minutes and then placed in CAS-Block (Thermo Fisher; catalog No. 008120). The following primary antibodies were used: insulin (Cell Signaling Technology; catalog No. 4590), glucagon (Millipore; catalog No. 4030-01F), VE-cadherin (vascular endothelial cadherin; Santa Cruz; catalog No. sc-6458), Pecam1 (Santa Cruz; catalog No. c-1506), Plvap (BD Pharmingen; catalog No. 550563), and endomucin (Santa Cruz; catalog No. sc-65495). All antibodies were used at a l:00 dilution, and slides were incubated in the primary antibody overnight at 4 °C. Secondary antibodies conjugated to AlexaFluor were used for visualization. For transfected HUVECs, cells were fixed with 4% paraformaldehyde for 1 hour at room temperature permeabilized with 0.1% Triton X-100 for 30 minutes at room temperature, blocked with 10% donkey serum (Sigma; No. S30-100ML) for l hour at room temperature. The following primary antibodies were used: GFP (green fluorescent protein; Thermo Fisher Scientific; catalog No. A10262) and NKX2-3 (European Monoclonal Antibodies Network). All antibodies were used at a l:200 dilution overnight at 4 °C. Please see the Major Resources Table in the Supplemental Material.

In Situ RNAscope HiPlex v2 Assay

Tissue cross sections were obtained from E18.5 mouse pancreatic tissue. Pregnant CD1 dams were euthanized via CO₂ asphyxiation, in accordance with IACUC procedures at UT Southwestern Medical Center. For the RNAscope HiPlex v2 assay (Advanced Cell Diagnostics), the tissue was processed according to the manufacturer’s instructions. Briefly, slides were fixed in 4% paraformaldehyde for 1 hour. Slides were then washed 2× in PBS and dehydrated to 100% ethanol. Slides were permeabilized with 12 minutes of protease III treatment. Target probes were applied at l× concentration and incubated at 40° for 2 hours. Amplification, fluorophore conjugation, and cleavage were done in accordance with the manufacturer’s protocol. After all probes were imaged, slides were then cleaved l more time, washed 3× in PBS for 5 minutes per wash. One hundred fifty microliters of CAS-Block was placed on each slide for l hour at room temperature, and then primary antibodies were added at a dilution of l:100. Slides were then incubated in the primary antibody overnight at 4 °C. Secondary antibodies conjugated to AlexaFluor were used for visualization.

In Situ Visualization

Images were taken on Nikon A1R confocal, a laser scanning system using the ×40 objective (resolution of 3.2 pixels/l µm). For RNA scope images, tile-scan images coupled with zstacks were taken to obtain large images of the pancreatic section. Maximum intensity projections were created in ImageJ, and then images were registered using the ACD Hiplex Registration software.

In Situ Image Quantification

Images were quantified in a single-blind manner. Ten representative images were taken from the endocrine and exocrine regions (based on staining of insulin/glucagon) of the capillary vessels based on size. Afterward, these images were analyzed on Fiji by a blinded researcher. RNA probe channels for Pecam1 and genes of interest were thresholded using the MaxEntropy option and adjusted accordingly to remove the background. Then, the regions of interest were outlined manually based on the Pecam1/endomucin-staining channel. Afterward, the area of the probe was quantified and divided by the area of the regions of interest to obtain normalized ratios of the RNA probe to the area of the vessel.

Transfection

Plasmids expressing the NKX2-3 protein tagged to GFP on the C terminus (NKX2-3-GFP) and GFP only were purchased commercially (OriGene; catalog No. RG206667 and PS100010). Plasmids were replicated in Top10 E. coli (New England Biolabs; catalog No. C3019H) and purified using the ZymoPURE II Plasmid Maxiprep Kit (Zymo Research; catalog No. D4202). HUVEC (human umbilical vein endothelial cells) passages 2 to 5 (ScienCell; catalog No. 8000) were cultured in Endothelial Cell Medium (ScienCell; catalog No. 1001) on T75 flasks coated with 0.1% gelatin in PBS. HUVECs were harvested with TryplE Express (Thermo Fisher Scientific; catalog No. 12604013), and 500 000 cells per condition (NKX2-3-GFP and GFP only) were transfected with 8 μg of plasmid using the HUVEC Nucleofector transfection kit (Lonza; catalog No. VPB-1002) according to the manufacturer’s protocol. Transfected HUVECs were plated into 0.1% gelatin-coated 6-well plates for RNA assays and 96-well plates for immunofluorescence staining. Media was changed after 4 to 6 hours and cells harvested 24 hours after transfection. Images were also taken after media change and before harvesting using the BioTek Lionheart FX Automated Microscope.

RT-qPCR

Transfected HUVECs were harvested with TryplE Express (Thermo Fisher Scientific; catalog No. 12604013), and RNA was extracted using the RNeasy Micro Kit (No. 74004). RNA concentrations were quantified and 1 µg was reverse transcribed to cDNA using the iScript cDNA Synthesis Kit (BioRad; catalog No. 1708890). qPCR (quantitative polymerase chain reaction) was conducted on the Roche Lightcycler 480 Instrument II using the SsoAdvanced Universal SYBR Green Supermix (catalog No. 1725271) with custom primers synthesized by Thermo Fisher Scientific (Major Resources Table in the Supplemental Material). log2FCs were calculated using the 2^−ΔΔCp method. ΔCp was calculated using the average Cp of ACTB and GAPDH housekeeping genes.

Statistical Analysis

The DE analysis was conducted using the Wilcoxon rank-sum test for EC atlases or MAST for the integrated atlas. Afterward, the P values obtained were adjusted using the Bonferroni correction method. For quantification of immunofluorescence staining and RNA probe binding, sample variances were tested for equality using the var.equal function in R. If variances were equal, then a 1-tailed 2-sample t test was conducted; if variances were unequal, then a Welch t test was conducted. For RT-qPCR, a 1-tailed Wilcoxon rank-sum test was conducted. To compare module scores between bulk RNAseq samples, paired 1-tailed t test was conducted. Fisher z test was conducted as follows,³⁰ after which 1-way ANOVA followed by Tukey’s HSD (honestly significant difference) post hoc test was conducted to determine the differences in means of z test statistics across samples.

Let r₁ and r₂ be the Spearman ρ coefficients for samples with sizes n₁ and n₂, respectively.

The z transformation for r₁ and r₂ and z₁ and z₂, respectively, can be calculated as follows:

$${\displaystyle \begin{array}{l}{\displaystyle z_1=0.5\times \frac{ln(1+r_1)}{ln(1-r_1)}}\end{array}}$$

$${\displaystyle \begin{array}{l}{\displaystyle z_2=0.5\times \frac{ln(1+r_2)}{ln(1-r_2)}}\end{array}}$$

The SE for the difference in Z scores (SE_zdiff) can be calculated as follows:

$${\displaystyle \begin{array}{l}{\displaystyle se{z}_{diff}=\sqrt{\frac{1.06}{(n_1-3)}+\frac{1.06}{(n_2-3)}}}\end{array}}$$

The test statistic for the difference in Z scores (z_test) can be calculated as follows:

$${\displaystyle \begin{array}{l}{\displaystyle z_{test}=\frac{z_1-z_2}{se{z}_{diff}}}\end{array}}$$

Results

Comparative Decontamination Strategy Reveals a Specialized Pancreatic EC-Enriched Signature Conserved Across Species and Developmental Stages

To generate a transcriptomic profile of pancreatic ECs, we have conducted a comparative analysis of ECs in multiorgan sn/scRNAseq atlases and integrated human pancreatic EC sn/scRNAseq data to dissect the intertissue and intratissue pancreatic EC heterogeneity, respectively (Figure S1).

Briefly, to define signature genes characteristic of pancreatic ECs, we subsetted and pooled all cells annotated as ECs from all available organs from the adult mouse Tabula Muris and fetal human Descartes multiorgan sn/scRNAseq atlases (Table 1). EC annotation was verified based on the expression of canonical EC markers CDH5 (VE-cadherin), PECAM1 (CD31), and KDR (VEGFR2). This generated an EC-only atlas for fetal human (Figure S2A through S2C) and adult mouse (Figure S2D through S2F), respectively. We then identified DEGs enriched in pancreatic ECs compared with ECs from all other tissues in each atlas (Figure S2C and S2F; Tables S1 and S2). Consistent with previous studies, a careful inspection of these DEGs revealed a significant number of acinar and endocrine genes (INS and GCG), likely resulting from ambient RNA contamination from surrounding cells.³¹

Table 1.

Summary of Data Sets Used in This Study

To identify EC-specific genes, we compared the DEGs enriched in pancreatic ECs to the DEGs of other cell types in the pancreas. We reasoned that any DEGs that are enriched in ECs compared with other cell types in the pancreas could not be contaminating genes from other populations, as the contaminating genes should have a higher expression in the cell type it originates from. So, we curated a set of EC-enriched DEGs from 4 pancreatic sn/scRNAseq data sets (Figures S3 and S4). For each data set, we identified the significantly enriched DEGs in the pancreatic vascular ECs against all other nonvascular pancreatic cells, such as β-cells or acinar cells excluding lymphatic ECs and perivascular cells, as the expression could overlap with that of ECs (Table S3; Figures S3C, S3F, S4C, and S4F). There was limited overlap in the EC genes discovered in the 4 data sets (Figure S5A). We combined all unique DEGs identified in the 4 data sets into 1 set, which we refer to as the curated set of EC genes.

The overlap between the curated set of EC genes and the pancreatic EC-enriched DEGs from the Descartes and Tabula Muris EC atlases identified 65 pancreatic EC-enriched signature genes (Figure 1A and 1B; Table 2; Tables S4 and S5). Comparing the expression of the genes across both atlases identified NKX2-3 alongside PLVAP, IGFBP3, RAMP3, COL15A1, SYNPO, and EHD4 (Figure 1C) as top signature genes. Analysis also demonstrated that a vast number of genes were potential contaminants (Figure S5C and S5D). We verified the pancreas specificity (Figure S6) and EC specificity (Figure S7) of these signature genes.

Figure 1.

A conserved pancreas endothelial cell (EC)–enriched gene signature exists across species and age. A through C, Uncovering biologically relevant pancreas EC-enriched signature genes. A, Pancreatic EC-enriched signature genes lie at the overlap between pancreatic enriched differentially expressed genes (DEGs) derived from EC atlases and EC-enriched DEGs derived from pancreatic data sets (curated set of EC genes). B, The overlap of significant pancreatic enriched DEGs in the Descartes fetal human EC atlas and Tabula Muris adult mouse EC atlas and the curated set of EC genes. C, The expression of pancreas EC-enriched signature genes across the 2 multiorgan EC atlases. The x and y axes demonstrate the average log2 fold change (log2FC) of the genes in pancreas ECs compared with other tissue ECs for each EC atlas. The normalized percentage difference was calculated by subtracting the percent expression in population 1 (pancreatic ECs) with that of population 2 (the expression in all other tissue ECs) and was weighted by dividing the difference by the percent expression in population 1. The normalized percentage differences are visualized in color scales and dot sizes for each EC atlas. D through F, Pancreas EC-enriched signatures genes were validated independently in the adult human Tabula Sapiens multiorgan atlas. D, Module scores derived by assessing the collective expression of the pancreas EC-enriched gene signature in the different tissue ECs. Each violin is colored based on the average scaled module score for that tissue. E, The individual log-normalized average scaled expression of the top 30 pancreas EC-enriched signature genes in the Tabula Sapiens EC atlas. Genes are ordered based on normalized percentage difference, and hierarchical clustering of tissues is based on Euclidian distances between the average scaled expression of all the signature genes weighted by the percentage expression. F, RNAScope of Nkx2-3, Plcb1, Clec14a, and Heg1 on embryonic day (E) 18.5 mouse pancreas tissue sections. Islet and exocrine regions are identified by immunofluorescence costaining of insulin and glucagon (pink), which stains cells of the islets; vasculature is identified by costaining of Pecam1 and endomucin (neon blue); epithelial cells are stained by E-cadherin (yellow); nuclei are stained by DAPI (4′,6-diamidino-2-phenylindole; blue); and genes of interest are identified by RNAScope probes (red).

Table 2.

EC-Enriched Signature Genes

Pancreatic EC-Enriched Signature Was Validated in an Independent Multiorgan Atlas

To further increase confidence in the specificity of the pancreatic EC-enriched signature, we validated it in an independent, multiorgan atlas, not used to derive this signature. For this purpose, we analyzed the recently published adult human Tabula Sapiens multiorgan sn/scRNAseq atlas. Its corresponding EC atlas was generated by subsetting all annotated ECs (Figure S8A and S8B), as done for other atlases above. We then conducted graph-based clustering on the EC atlas using only the expression data from the signature genes. This identified cluster 4 that is made up of only pancreatic ECs, and vice versa, pancreatic ECs are made up of majorly cluster 4 (Figure S8C and S8D). This demonstrated that expression data from only pancreatic EC signature genes are sufficient to classify pancreatic ECs from other tissue ECs. This clustering was driven by the genes EHD4, NKX2-3, PLVAP, COL15A1, SYNPO, and SPRY1 (Figure S8E). Collective module score¹⁹ of the signature gene expression demonstrated enrichment of the signature genes within the pancreatic ECs of the Tabula Sapiens EC atlas (Figure 1D). Hierarchical clustering based on the average expression and percent expression of signature genes also allowed pancreatic ECs to cluster in a separate branch (Figure 1E). We have also confirmed the vascular expression of Nkx2-3 and other enriched genes (Plcb1, Clec14a, and Heg1) in situ in embryonic day (E) 18.5 mouse pancreas (Figure 1F).

Pathways Driving EC-Enriched Gene Signatures Were Identified by Gene Set Overrepresentation Analysis

We then conducted gene set overrepresentation analysis to identify pathways driving the pancreatic EC-enriched signature (Figure 2; Table S6). This implicated key signaling pathways including TGF-β (transforming growth factor-β), RHO/RAC GTPase (guanosine triphosphatase), PI3K/AKT (phosphoinositide 3-kinase/protein kinase B), and PDGF (platelet-derived growth factor) signaling. Interestingly, several signature genes are also involved in S1P (sphingosine-1-phosphate)-mediated regulation of EC barrier integrity via cytoskeletal alterations regulated by the RHO/RAC GTPase signaling pathway such as AHR,³² RDX,³³ MSN,³³ SYNPO,³⁴ RFNLB, and SPNS2.³⁵ As such, we also assessed other genes involved in S1P signaling such as sphingosine kinases, phospholipid lipases, and S1P receptors in the pancreatic data set and across ECs in the adult human atlas (Tabula Sapiens; Figure S9). We observed that pancreatic ECs had the highest expression of S1PR1 and PLPP1 compared with other pancreatic cell types, including β-cells, which did not notably express any of these genes directly. PLPP1 was also most enriched in pancreatic ECs compared with other tissue ECs.

Figure 2.

Gene set overrepresentation analysis of pancreas endothelial cell (EC)–enriched signature genes. Gene set overrepresentation analysis was conducted on the signature genes using Enrichr (v3.2) with the following pathway databases: WikiPathway (2023), Reactome (2022), KEGG (Kyoto Encylopedia of Genes and Genomes; 2021), BioPlanet (2019), Panther (2016), and BioCarta (2016). Selected pathways of interest are visualized in a network; genes are shown in blue ellipses and terms in boxes; boxes are colored based on the negative log-adjusted P value from the gene set overrepresentation analysis.

Comparison Across Multiorgan Atlases Also Identified Pancreatic EC-Enriched Genes Unique to Species and Age

To ensure we did not overlook any biologically relevant information, we further identified all pancreatic EC-enriched DEGs in the adult human EC atlas (Tabula Sapiens). Similar to the other 2 EC atlases, the top pancreatic EC-enriched DEGs were contaminating genes (Figure S10A through S10C; Table S7). As before, we filtered contaminating genes using the curated set of EC genes (Table S8). We then compared these DEGs with the filtered pancreatic EC-enriched DEGs from the fetal human (Descartes) and adult mouse (Tabula Muris) EC atlases. Overall, 34 pancreatic EC-enriched DEGs overlapped across all 3 atlases (Figure S10B). Beyond the conserved DEGs across all 3 atlases, DEGs that are conserved across only 2 atlases are also relevant as they can represent human but not mouse-specific genes or genes relevant in adult but not fetal development. The analysis identified CD320, FAM13C, VWA1, SOX18, ACE, GPRC5B, IGFBP4, CDC42EP1, CCDC85B, and TMEM88 as the top genes unique to humans, that is, excluding Tabula Muris (Figure S10D), and FAM167B, CCN3, HOXB5, SNAI2, LAMB1, CYSLTR1, CD24, and JAM3 as the top genes unique to adults, that is, excluding Descartes (Figure S10E).

Integrated EC Atlas Revealed the Pancreatic Vascular Hierarchy and Capillary Subpopulations

Next, we assessed the intratissue EC heterogeneity in the pancreas. To increase the transcriptomic resolution, we used Harmony²¹ to integrate the pancreatic ECs across 3 human data sets. This produced an atlas of human pancreatic ECs with 6 different subpopulations (Figure 3A) that were present in similar proportions across the individual data sets (Figure 3B and 3C). These populations corresponded to the topographical location of ECs along the vascular tree. Three arterial populations were identified by the expression of GJA5,³⁶ GJA4,³⁷ and EFNB2.³⁸ Large artery ECs also expressed ELN and SULF1, which are involved in artery ECM (extracellular matrix) formation.³⁹ Two capillary populations were annotated based on CD36 expression, which has also been shown to be enriched in capillary ECs in the lung.⁴ One of the arterial populations coexpressed CD36 and was annotated as a precapillary arteriole population and the remaining annotated as a feeding arteriole population. We also annotated 1 postcapillary venule population based on the expression of the venule marker DARC⁴⁰ and capillary marker CD36 (Figure 3D).

Figure 3.

Integrated atlas of human pancreatic endothelial cells (ECs) identifies specialized capillary subpopulation. Three different human pancreatic EC data were subsetted from their parent data and integrated using Harmony. This includes the fetal human pancreas data set from the Descartes gene expression atlas and the neonatal and adult pancreas data set from Tosti et al. A, The resulting integrated uniform manifold approximation and projection (UMAP). B, The percentage distribution of different subpopulations in each data set. C, The UMAP of the integrated map split by data set. D, The expression of markers used to differentiate the vascular bed subclusters in the integrated EC atlas. E, The expression of selected genes across biological processes of interest in the atlas. F, The UMAP of the integrated atlas colored by pseudotime inferred through trajectory analysis using SCORPIUS. The trajectory is colored in blue. G, The heatmap demonstrating the change in average scaled expression of the top unique 50 markers determined by Wilcoxon rank-sum test for each subpopulation across pseudotime. Genes are hierarchically clustered based on the Euclidian distance between the average scaled expression of the genes. H, The expression of the top 12 genes across pseudotime ranked by importance to trajectory inference using SCORPIUS.

The DE analysis (Table S9) across the different subpopulations revealed genes involved in caveolae and fenestrae formation, ECM production, IGF (insulin-like growth factor) signaling, SOXF (SRY-related HMG-box subgroup F) TFs, immune activity, and membrane transport. Therefore, we assessed other genes in these families to identify potential trends that correspond to vascular hierarchy (Figure 3E). We showed that capillary 1 had higher expression of several caveolae- and fenestrae-associated genes such as PLVAP,⁴¹ SDPR (or CAVIN2),⁴² and EHD4.⁴³ It also had a higher expression of several immune-related genes, including FABP5,⁴⁴ CD74,⁴⁵ ITM2A, B2M, HLA-C, and HLA-B. Higher immunologic activity and presence of fenestrae are characteristics that have previously been attributed in the literature to intraislet capillaries in comparison to exocrine capillaries.⁴⁶ Capillary 1 was also highly enriched for markers identified for islet ECs in sn/scRNAseq studies of isolated islets (Figure S11A and S11B). We also conducted annotation transfer mapping using the integrated atlas as a reference to annotate the Tabula Sapiens pancreatic ECs, which was not used to derive this atlas (Figure S11C and S11D). Altogether, capillary 1 and capillary 2 subpopulations were annotated as islet and exocrine capillary ECs, respectively. Other genes highly enriched in islet capillary ECs included SPARC, SPARCL1, IGFBP7, SLCO2A1, CD320, and SLC2A3. By contrast, exocrine capillary ECs displayed a more quiescent expression pattern with few significantly enriched genes, which included AKAP12, PIEZO2, and NEURL1B. The precapillary arteriole population also had an interesting gene expression pattern, with high expression of HEG1, SOX7, IGFBP3, and INSR.

Pseudotime trajectory inference (Figure 3F) predicted a trajectory from the venule population to the artery populations, through the capillary populations. Gene expression of top markers for each vascular bed subpopulation across pseudotime also demonstrated a continuous gene expression pattern across the vascular topography rather than discrete EC subpopulations (Figure 3G). We also identified genes that predict the order of the cells along pseudotime (Figure 3H). This allowed us to identify the expression pattern of both established (GJA4, GJA5, and DARC) and understudied arteriovenous markers for differentiating ECs along the vascular tree. Of these genes, we found that capillaries have peak expression of PLVAP and APLNR. Another gene of interest is UNC5B, which was highly enriched in arteriole ECs. Several other genes (POSTN, GPR126, and ZNF385D) match the expression pattern of DARC, which peaks specifically in the postcapillary venule ECs.

Islet Capillary ECs Are Driving the Pancreatic EC-Enriched Signature

We then assessed the pancreatic EC-enriched gene signature in the integrated atlas. This demonstrated that the pancreatic EC specialization is driven by islet capillary ECs followed by the precapillary arteriole population (Figure 4A). We also analyzed a bulk RNAseq data set published by Jonsson et al,⁴⁷ who isolated ECs from islet and exocrine pancreas from 8 adult human donors. Preliminary PCA (principal component analysis) demonstrated that the islet sample for donor 8 is an outlier (Figure S12); hence, donor 8 sample was removed before the downstream analysis. We then assessed the pancreatic EC-enriched signature in the remaining exocrine and islet EC samples in the bulk RNAseq data set and observed a significant enrichment within islet EC samples compared with exocrine EC samples (Figure 4B). We also conducted PCA with only the pancreatic EC-enriched signature gene expression to demonstrate that it is sufficient to classify exocrine and islet EC samples (Figure 4C). Assessing the genes driving the PCA highlighted that NKX2-3, PLVAP, and COL4A1 were enriched in islet EC samples, while MEOX1, PCDH17, and TNFSF10 were enriched in exocrine samples (Figure 4D).

Figure 4.

Pancreas endothelial cell (EC)–enriched gene signature is enriched in islet capillary ECs. A, The module scores of the pancreas EC-enriched gene signature in the different EC subpopulations in the integrated EC atlas. B, Box plot of the collective module score for the pancreas EC-enriched signature in bulk RNA sequencing (RNAseq) islet and exocrine ECs from healthy normoglycemic human donors sequenced by Jonsson et al. Each dot represents 1 biological replicate, defined as 1 donor (n=7). The P value from a paired t test is shown. C, The principal component analysis of the islet and exocrine bulk RNAseq EC samples using only log normalization of counts per million (logCPM) scaled expression values of the signature genes. D, The heatmap of the average scaled expression of the signature genes with samples ordered based on PC1 (principal component 1) scores and genes ordered based on PC1 feature loading scores. E, The individual dotplot of selected genes showing the average scaled expression in the different subpopulations in the integrated atlas. F, Box plots of the same selected genes with the logCPM values of the bulk RNAseq islet and exocrine EC samples. Adjusted P values from the DE analysis of the bulk RNAseq data using DESeq2 are shown. Each N represents 1 donor (n=7).

We further conducted the DE analysis (Table S10) to identify all significant DEGs in islet and exocrine samples in the bulk RNAseq data. Of the signature genes, IGFBP3, SOX4, CCND1, MCAM, COL4A1, NKX2-3, HSPG2, TBC1D1, FMNL3, ITGA5, VAT1, COL4A2, PLVAP, SNTB2, SYNPO, and GRN were significantly enriched in islet EC samples and 3 genes were significantly enriched in exocrine samples: PCDH17, MEOX1, and TNFSF10.

Importantly, the bulk RNAseq DE analysis aligned with the integrated atlas expression patterns (Figure 4E and 4F; Figure S13). Islet EC bulk RNAseq samples were significantly enriched for SPARC, NKX2-3, PLVAP, VAT1, and CCND1, which were enriched in islet capillary ECs in the integrated atlas. Islet EC samples were also significantly enriched for COL4A1, COL4A2, IGFBP3, MCAM, SOX4, FMNL3, and INSR, which were enriched in the precapillary arteriole ECs in the integrated atlas. Exocrine EC samples were significantly enriched for PCDH17 and NEURL1B, which were enriched in exocrine capillary ECs in the integrated atlas. Exocrine samples were also significantly enriched for MEOX1 and TNFSF10, which were enriched in postcapillary venule and large artery ECs, respectively, in the integrated atlas. This suggests that the islet samples in the bulk RNAseq data consist of ECs characterized as precapillary arterioles and intraislet capillary ECs, and the exocrine samples consist of exocrine capillary ECs, large artery ECs, and venule ECs.

Capillary Subpopulations Were Also Identified by Spatial Transcriptomics

To confirm the spatial localization of the capillary subpopulations, we established unique signature genes for each subpopulation from the integrated atlas (Figure S14 and S15) and assessed the signatures in a spatial transcriptomics data from Olaniru et al,⁴⁸ who sequenced fetal pancreatic tissue samples from postconception weeks 12 to 20 using the Visium platform. Olaniru et al have deconvoluted their spatial data and provide cell-type proportions for each voxel in the samples. Using this, we can determine the proportions of islet cells (α, β, δ, and endocrine progenitors) and exocrine cells (acinar and ductal) across the samples (Figure S16A). Samples from only postconception weeks 18 and 20 were analyzed as they demonstrated clear distinction in exocrine and islet regions (Figure S16B).

Subsequently, we calculated scores for the different vascular subpopulation signatures in the spatial transcriptomic samples and assessed its association with islet and exocrine proportions (Figure S17). The analysis revealed that the islet capillary signature had the highest Fisher z test score, indicating an enrichment in islets over exocrine pancreas (Figure 5A). In contrast, the exocrine capillary signature showed the lowest score, a negative score, suggesting enrichment in the exocrine regions. These data demonstrate that there exists an exocrine capillary EC subpopulation in the exocrine pancreas that is distinct from the islet capillary EC subpopulation within the islets (Figure 5B and 5C).

Figure 5.

Specialized capillary population is located within the islets. Signature genes for vascular subpopulations identified from the integrated atlas were assessed on spatial transcriptomics data from the study by Olaniru et al. Spearman correlations were calculated for the signature scores with deconvoluted islet and exocrine proportions. A, The Fisher Z score of the difference between Spearman correlation of the signature with islet proportions and exocrine proportions. Each N represents 1 biological replicate, defined as 1 tissue sample sequenced (n=4). The line inside each box represents the median, with the box edges indicating the interquartile range (IQR). Whiskers extend to 1.5× the IQR, and outliers are shown as points beyond the whiskers. P values were obtained using ANOVA, followed by Tukey’s HSD (honestly significant difference) post hoc test for pairwise comparisons. B and C, Islet capillary and exocrine capillary scores in mutually exclusive islet and exocrine regions for postconception week 20 sample 2, characterized as >10% of cells in each voxel identified as islet cells and 0% identified as exocrine or vice versa. B shows the corresponding voxels identified as islet (pink) and exocrine (purple). C visualizes the signature scores for islet capillary and exocrine capillary endothelial cells (ECs) in exclusively islet and exocrine voxels. Arrows indicate selected regions that are different between capillary islet and capillary exocrine. D through H, In situ validation of islet EC capillary population in embryonic day (E) 18.5 mouse pancreas. D, The expression of 3 different DEGs that were enriched in the islet capillary population in the integrated pancreas EC atlas. These genes include PLVAP (left), NKX2-3 (middle), and SPARC (right). E, Immunofluorescence (IF) staining of Plvap. F and G, RNAScope images of Sparc (F) and Nkx2-3 (G). For the in situ images, pink shows IF costaining of glucagon and insulin, which highlights the islets from the exocrine tissue for all images; green shows PLVAP IF staining; red shows Sparc and Nkx2-3 RNAscope punctae; white shows Pecam1/VE-cadherin (vascular endothelial cadherin) costaining; and neon shows Pecam1/endomucin costaining to highlight vessels; yellow shows E-cadherin staining. D, Quantification of fluorescence intensity in arbitrary units (AU) for Plvap (left) and RNAscope area of probe binding (μm²) in islet and exocrine capillaries for Sparc (middle) and Nkx2-3 (right). Each N represents 1 biological replicate, defined as 1 vessel (n=59 for islet and n=35 for exocrine regions for Plvap IF, and n=10 for Sparc and Nkx2-3 RNAscope). The line inside each box represents the median, with the box edges indicating the IQR. Whiskers extend to 1.5× the IQR, and outliers are shown as points beyond the whiskers.

We further corroborate the spatial data by conducting RNAscope and immunofluorescence staining of selected islet capillary genes identified in the atlas, in situ on E18.5 mouse pancreatic tissue (Figure 5D and 5H). We assessed 3 genes identified to be enriched in islet capillary ECs: PLVAP, NKX2-3, and SPARC (Figure 5D). RNAscope data demonstrated that a substantial portion of the probes, 70.1% (±3.44%) for Sparc and 49.6% (±5.44%) for Nkx2-3, was colocalized within the vessels (Figure S18). Immunofluorescence staining and quantification of Plvap showed that it had over 2× higher fluorescence intensity (3970±571 AU) in islet Pecam1/Emcn⁺ capillaries compared with exocrine capillaries (1513±268 AU; Figure 5E and 5H). Further, Sparc in the islet vessels had over 3× as much RNA expression level as measured by area of bound probe (0.0416±0.0086 µm²) compared with the exocrine vessels (0.01234±0.0012 µm²; Figure 5F and 5H). RNAscope of Nkx2-3 did not show any statistically significant difference in islet versus exocrine vessels (P=0.08; Figure 5G and 5H).

Motif Analysis of NKX2-3 Binding and In Vitro Expression Assays Demonstrated That NKX2-3 Is a Key Mediator of Pancreatic EC Identity

NKX2-3 was a key pancreas EC-enriched signature gene that was also enriched in islet capillary ECs in the integrated atlas and bulk RNAseq data of isolated islet EC samples. Since NKX2-3 is a key developmental TF,⁴⁹ we sought to determine whether NKX2-3 could be involved in modulating the expression of the other pancreatic EC signature genes. TF motif analysis (Figure 6A and 6B) using the HOMER software²⁶ to assess potential NKX2-3 DNA-binding motifs⁵⁰ revealed that the NKX2-3 motif was present within close proximity (+200 to −1000 bp) of the transcriptional start site of 37% of the pancreatic EC-enriched signature genes (24/65). These included PLVAP, HEG1, AHR, SYNPO, COL4A1, PCDH17, and HSPG2 (Figure 6B).

Figure 6.

NKX2-3 regulates pancreas endothelial cell (EC) gene expression. A, TF (transcription factor) motif analysis was conducted to find the NKX2-3 DNA-binding motif obtained from JASPER, an open-access database for TF binding profiles (MA0672.1), within +200 to −1000 bp of transcription start site (TSS) of pancreas EC-enriched signature genes. B, Twenty-four signature genes identified with the NKX2-3 motif are shown. C and D, NKX2-3 tagged C terminally to GFP (green fluorescent protein; NKX2-3-GFP) and GFP only was induced transiently in HUVECs (human umbilical vein endothelial cells). C, Immunofluorescence image of HUVECs transfected 24 hours prior with an NKX2-3-GFP at ×40 magnification. NKX2-3 is shown in red, GFP in green, and DAPI (4′,6-diamidino-2-phenylindole) in blue. D, Gene expression was assessed by RT-qPCR (reverse transcription quantitative polymerase chain reaction) 24 hours after transfection. The mean log2 fold change over GFP-only induced HUVEC genes with P<0.05 based on 1-tailed Wilcoxon rank-sum test is shown. Each dot presents 1 biological replicate (n), defined as the result from RT-qPCR conducted on 1 transfection experiment. The error bars represent the SEM (n=5 for PLVAP; n=4 for SMAD7, HSPG2, CLEC14A, and EDN1; and n=3 for SPARCL1).

To define NKX2-3 TF’s potential role in regulating pancreatic EC gene expression and organotypic specialization, we transiently induced NKX2-3 tagged C terminally to GFP (green fluorescent protein; NKX2-3-GFP) in HUVECs in vitro and assessed the expression of selected signature genes. Transfection efficiency assessed by GFP expression was ≈80%, 4 hours after transfection, and ≈20%, 24 hours after transfection (Figure S19A and S19B). NKX2-3-GFP but not GFP only (control) was localized to the nucleus (Figure S19A and S19B; Figure 6C). RT-qPCR (reverse transcription quantitative polymerase chain reaction) conducted on HUVECs 24 hours post-transfection (Figure 6D; Figure S19C) demonstrated that NKX2-3-GFP–induced HUVECs had significantly increased PLVAP, SPARCL1, SMAD7, and EDN1 expression and significantly reduced CLEC14A and HSPG2 expression compared with GFP-only–induced HUVECs (Figure 6D).

Discussion

ECs demonstrate a remarkable extent of intratissue and intertissue heterogeneity in the structure, function, and molecular characteristics. Omics technologies now enable us to dissect this organotypic heterogeneity down to the single-cell resolution. Here, we have analyzed different multiorgan gene expression atlases and have identified a gene signature for pancreatic ECs across species and age. These genes are highly conserved and potentially relevant, as they were identified as noncontaminant EC genes that were significantly enriched in the pancreas across 2 different EC atlases and validated independently on a third EC atlas, which were generated from different species (human and mouse), sequencing methodologies (10× and smartseq2), and sample age (adult and fetal). Furthermore, we have established an integrated atlas of human pancreatic ECs using 3 recently published data sets from fetal, adult, and neonatal samples. This atlas allowed us to identify distinct subpopulations with unique gene expression patterns corresponding to the topographical location of ECs along the vascular tree. Specifically, we describe 2 capillary subpopulations: a specialized intraislet capillary and a less-specialized exocrine capillary. Importantly, the analysis identified a novel role for the TF NKX2-3 in regulating the expression of PLVAP and other pancreatic EC-enriched signature genes.

The importance of profiling the transcriptomics of pancreatic ECs is clear when considering their integral role in β-cell development and function and its relevance in diabetes.^51–53 The pancreatic EC-enriched signature genes can shed insight into the development of specialization in glucose homeostasis and insulin transport within the islet microvasculature. Furthermore, owing to the tissue-specific roles of ECs, there is a recent push toward engineering physiologically relevant 3-dimensional vascularized tissues with tissue-specific characteristics.^54,55 However, past attempts at generating a transcriptomic signature of pancreatic ECs have inadvertently introduced bias from contaminating/ambient RNA from acinar cells.^7,8 We circumvent this issue through a comparative analysis framework of DEGs across multiple atlases and pancreatic data sets. Importantly, this framework can be ported to identify a biologically relevant tissue-specific signature for any cell type of interest.

One of the top genes of interest across all 3 EC atlases analyzed in this study was NKX2-3. NKX2-3 also had a higher expression in islet capillary ECs compared with exocrine capillary ECs in the integrated human pancreatic EC atlas and bulk RNAseq data. NKX2-3 DNA-binding motif analysis demonstrated a potential regulatory role in ≈40% of the pancreatic EC-enriched signature genes. We demonstrated that induction of NKX2-3 in HUVECs can increase PLVAP expression. PLVAP is an established marker of fenestration,⁴¹ whose regulation is still poorly understood. In this study, we also demonstrated PLVAP as another top pancreas EC-enriched signature gene. Other signature genes increased by the induction of NKX2-3 are SPARCL1, SMAD7, and EDN1. This is in line with the gene set overrepresentation analysis that identified TGF-β signaling is enriched in pancreatic ECs, as SMAD7 is an established regulator of TGF-β signaling.^56,57 The causal link between NKX2-3 and the expression of other pancreatic signature genes suggests a potential role for NKX2-3 as a regulator of pancreatic EC identity and function. However, not all signature genes analyzed were upregulated by NKX2-3 expression. This could be due to several factors such as the low efficiency of transfection, which could be masking the true effect of NKX2-3 induction, the requirement of coactivators for these genes, or the lack of a direct role for NKX2-3 in the expression of some of the signature genes.

NKX2-3 is a member of the NK2 family of homeobox domain-containing TFs. These are critical organ-specific developmental TFs.^49,58 NKX2-3 is involved in the development of the spleen and gut,⁵⁹ both of which arise from the endoderm, which also gives rise to the pancreas. Specifically, NKX2-3 has been shown to be integral for proper vascular specification in the gut and spleen organs.^60,61 However, NKX2-3 has not been previously implicated in vascular specification of the pancreas. Primarily, NKX2-3 is involved in regulating the expression of MAdCAM-1 (mucosal vascular addressing cell adhesion molecule 1).^62,63 Beyond MAdCAM-1, NKX2-3 has also been shown to regulate the expression of other important angiocrine factors such as EDN1, which is in line with the RT-qPCR data.⁶⁴ The murine EC atlas has also described Nkx2-3 as a tissue-specific marker for small intestine, colon, and spleen ECs.⁶ This makes NKX2-3 TF an exciting target for further studies in its role in islet EC identity.⁶⁵

In addition, we highlight several key transcriptomic differences that align with existing literature on islet EC function, paving the way for more in-depth studies. Of note, we identify several genes of interest involved in S1P signaling. S1P is a bioactive lipid, which is secreted by platelets, red blood cells, and ECs,⁶⁶ that has garnered much interest due to its multifaceted role in embryonic development, organ homeostasis, and disease pathogenesis.⁶⁷ In the pancreas, vascular-derived S1P is involved in pancreatic development.⁶⁸ Further evidence suggests that extracellular S1P may induce glucose-stimulated insulin secretion in β-cells.^69,70 In fact, exogenous S1P has been shown to improve β-cell proliferation in a type 2 diabetic mouse model.⁷¹ In this study, we highlight several genes involved in S1P signaling. Notably, S1P receptor genes (S1PR1–5) were not detected in β-cells across all pancreatic datasets, whereas pancreatic ECs highly expressed S1PR1 compared with other pancreatic cell types. Thus, it is likely that S1P signaling primarily involves the islet ECs, which then propagates its effects to β-cells through other means to regulate their function. Further research is needed to fully elucidate the role of S1P and its associated proteins in the β-cell/EC axis.

Finally, we identified a precapillary arteriole population enriched for COL4A1 and COL4A2, IGFBP3, INSR, and PDGFB compared with other EC populations in the integrated atlas. The precapillary arteriole signature was also highly correlated with islet regions in the spatial data. This indicates potentially important roles of the precapillary arterioles in insulin signaling in the pancreas and is in line with a previously reported role for insulin signaling in promoting vasodilation via NO.⁷² Recently, it was established that pericytes, which are characterized by PDGFRB (platelet-derived growth factor receptor β),⁷³ can also regulate blood flow in islet capillaries in response to changes in blood glucose levels.⁷⁴ Therefore, a mechanism whereby precapillary arterioles are regulating blood flow with cross talk with pericytes through interstitial insulin signaling and PDGFRB signaling from β-cells is likely.

Furthermore, islet capillary ECs also had higher expression of the vitamin B₁₂ transporter CD320, prostaglandin transporter SLCO2A1, glucose transporter SLC2A3 (or GLUT3), IGFBP7, and COL15A1. Therefore, their roles should be further investigated.

This study has limitations. The in situ validation was not performed on adult pancreas, and, except for Plvap, we did not conduct extensive protein-level validation. This is because of the high levels of hydrolytic enzymes in the exocrine pancreas, which we found caused both tissue and RNA probe degradation, as well as autofluorescence, leading to nonspecific background staining that interfered with the accuracy and sensitivity of staining. We also demonstrated the potential regulatory function of NKX2-3 through transient transfection of HUVECs. A stable NKX2-3–expressing EC line would allow for functional characterization over longer time periods.

Despite these limitations, we have succeeded in removing the noise that is inherent in pancreatic EC sn/scRNAseq data sets to uncover the transcriptomic profile of pancreatic ECs. Genes of interest and transcriptomic differences highlighted in this study indicate important functional roles of pancreatic ECs in glucose homeostasis and diabetes pathophysiology and identify important TFs that can aid future endeavors at generating pancreatic islet ECs for application in in vitro models. We have also generated an online single-cell resource that can be further explored by the community via the R Shiny application (app).

Article Information

Acknowledgments

S.T. Khan and S.S. Nunes conceptualized the study. S.S. Nunes obtained funding for the study. S.T. Khan curated, processed, analyzed, and visualized all single-nuclei/single-cell RNA-sequencing (RNAseq) data and spatial transcriptomics data. S.T. Khan and S. Vohra analyzed and visualized the bulk RNAseq data. S. Vohra and S.T. Khan generated the R Shiny application (app). N. Ahuja and O. Cleaver performed sample collection, staining, imaging, and acquisition of RNAscope and immunofluorescence (IF) data from mice. S. Taïb quantified all RNAscope data. S.T. Khan conducted in vitro assays. S.S. Nunes supervised the study. S.T. Khan and S.S. Nunes wrote the original manuscript. S. Taïb, S. Vohra, N. Ahuja, and O. Cleaver were involved in reviewing and editing the manuscript. We would also like to acknowledge the team of Dr Giovanna Roncador from the Centro Nacional de Investigaciones Oncológicas and the European Monoclonal Antibodies Network for providing the NKX2-3 (NK2 homeobox family 2 transcription factor) monoclonal antibody. The graphic abstract was made using Biorender.com.

Sources of Funding

This study received funding from the Canadian Institutes of Health Research (CIHR)–Juvenile Diabetes Research Foundation (JDRF), team grant (CIHR–ASD 173662, JDRF–5 SRA 2020 1058) to S.S. Nunes, Canada Foundation for Innovation, John R. Evans Leaders Fund (39909) to S.S. Nunes, Canada First Research Excellence Fund, Medicine by Design (MBD), Team Project Award (MBD Cycle 2) to S.S. Nunes, Ministry of Research, Innovation and Science, Early Researcher Award (ER17 13 14) to S.S. Nunes, Natural Sciences and Engineering Research Council of Canada (NSERC), Discovery Grant (RGPIN-2017-06621) to S.S. Nunes, and National Institutes of Health (HL113498 and DK121408) to O. Cleaver. S.T. Khan was partially supported by funding from the NSERC Collaborative Research and Training Experience Program (training program in organ-on-a-chip engineering and entrepreneurship) and the Thomas Zachos Scholars (the Mitochondrial Innovation Initiative). N. Ahuja received postdoctoral fellowships from JDRF (3-pdf-2023-1327-AN). S. Taïb received postdoctoral fellowships from the Banting and Best Diabetes Centre and the Toronto General Hospital Research Institute.

Disclosures

None.

Supplemental Material

Tables S1–S10

Figures S1–S19

Major Resources Table

Supplementary Material

atv-44-2596-s001.pdf ^{(11.3MB, pdf)}