We identified signature markers, transcriptional networks, angiocrine signaling pathways, and cellular subpopulations enriched in endothelial cells from various mouse tissues.
We uncovered sex differences in tissue-specific endothelial gene expression.
We found that markers of tissue-specific endothelial cells are conserved between mice and humans.
Novel endothelial cell membrane surface markers can be targeted for tissue-specific drug delivery.
Differentially expressed genes between male and female tissue-specific endothelial cells can be exploited to develop sex-specific cardiovascular disease models and treatments.
Endothelial cells (ECs) comprise the innermost lining of blood and lymphatic vessels. ECs play a critical role in tissue homeostasis by regulating blood flow, delivering plasma-borne macromolecules, assisting vessel formation, and contributing to adhesion of circulating blood cells.1 As such, EC dysfunction can contribute to various disease mechanisms, including atherosclerosis and coronary artery disease,2 tumor vascularization,3 diabetic complications,4,5 and neurodegenerative disease.6
Whereas ECs are considered a single cell type, they exhibit considerable structural, phenotypic, and functional heterogeneity depending on the tissue in which they reside.7–10 Tissue-specific EC dysfunction can contribute to a number of different diseases. In the blood–brain barrier, for example, ECs are bound by tight junctions to maintain a highly selective, low-permeability barrier. Endothelial dysfunction in the blood–brain barrier can lead to Alzheimer disease, epilepsy, and multiple sclerosis.11–13 The cardiac endothelium plays a crucial role in promoting cardiomyocyte proliferation and maturation via paracrine signaling,14–16 whereas limited EC proliferative potential results in suboptimal repair of damaged heart tissue after ischemic injury.17 Secretion of angiocrine factors from pulmonary ECs has been shown to improve lung alveolar regeneration,18 whereas angiocrine factors from liver sinusoidal ECs are critical in modulating hepatic regeneration.19 Thus, understanding tissue-specific EC functionality is critical for treating a wide range of human diseases.
Given that ECs in different tissues have unique functions, it is reasonable to hypothesize that they also exhibit unique molecular signatures. Indeed, bulk microarray-based transcriptomic analyses of mouse20,21 and human ECs,22 as well as bulk RNA sequencing of human fetal ECs,23 have demonstrated that ECs from different tissues have unique gene expression profiles. Bulk measurements, however, are unable to resolve transcriptomic heterogeneity that may exist between ECs from a given tissue. In addition to ECs from different vessel types (ie, artery, vein), there may be unidentified subpopulations of ECs within a tissue.
The recent advent of single-cell RNA sequencing (scRNA-seq) addresses this limitation by allowing in-depth transcriptomic analysis of thousands of single cells at unprecedented resolution.24 Recent studies have compiled single-cell transcriptomic25,26 and epigenomic27 data from all major mouse organs, including the endothelium.10 Notably, the Tabula Muris investigators constructed an organism-wide transcriptomic profile using both plate-based and microdroplet-based scRNA-seq, resulting in an analysis of over 100 000 cells from 20 murine organs and tissues.26 From this dataset, we extracted individual EC transcriptomes from 12 organs to investigate the transcriptomic landscape of tissue-specific ECs. Using these transcripts, we determined markers, signaling pathways, and biological processes enriched in tissue-specific ECs. We found that markers of murine tissue–specific ECs were conserved in human fetal ECs. We used an unsupervised clustering approach to identify novel EC subtypes, and inferred potential paracrine EC–to–parenchymal cell interactions and sex differences in EC gene expression. We observed robust expression of Lars2 in ECs of male mice, but its expression is low and variable in female ECs. Taken together, we have harnessed scRNA-seq to delineate the transcriptomic heterogeneity that underlies tissue-specific EC function.
Analytic methods and the resulting data have been made available to other researchers for purposes of reproducing the results presented. Scripts used in the study are available on GitHub (https://github.com/Lei-Tian/multi-Organ-EC). Analyzed data in R objects are available on Figshare (https://figshare.com/articles/EC_TSNE_Robj/12170358).
We obtained Smart-Seq2 RNA sequencing libraries of fluorescence-activated cell sorted single cells directly from the Tabula Muris database.26 Seurat objects generated by the Tabula Muris investigators were downloaded from Figshare (https://figshare.com/account/home#/projects/27733). The majority of scRNA-seq data analysis was performed using the Seurat R package.28 Annotations of each cell were downloaded from the Tabula Muris Github website (https://github.com/czbiohub/tabula-muris). ECs in all possible organs (12 in total) were extracted from the raw data. Cells with <500 detected genes and <50 000 confidentially mapped reads were excluded from downstream analysis. Raw counts were converted to log counts per million by log-normalization and subsequently scaled. The effects of confounding factors, including ribosomal RNA, total number of reads, and percentage of External RNA Controls Consortium controls, were removed by linear regression. We then selected 5203 highly variable genes (average expression ≥0.1 and dispersion ≥0.5) for downstream analysis.
Microarray-based transcriptomic measurements (Affymetrix GeneChip Mouse Gene 1.0 ST Array) of tissue-specific murine ECs were acquired previously by Nolan and colleagues20 and accessed from GEO Omnibus (GSE47067). All microarray data were processed using the oligo and limma R packages.29,30 Raw intensity values were background-corrected, normalized, and summarized using the robust multiarray average algorithm. Microarray data were compared with scRNA-seq data by averaging gene expression values across all replicate samples or cells. The top 10 differentially expressed genes (DEGs) in tissue-specific EC microarray data were identified using the empirical Bayesian method.28,31
To initially reduce the dimensionality of the scRNA-seq dataset, we performed principal component (PC) analysis on highly variable genes. We chose the top 20 PCs for downstream analysis based on a resampling procedure29 and the contribution of each PC to variance in the dataset. We then used the top 20 PCs to project cells onto 2-dimensional maps using both t-SNE (t-distributed stochastic neighbor embedding) and uniform manifold approximation and projection algorithms.32 Unsupervised clustering of single cells was performed using the FindClusters function in Seurat with the resolution measure set at 0.8. Briefly, this function first finds the k-nearest neighbors for each cell in the PC space. Then, connections between cells are weighted based on their Jaccard similarity, or the number of k-nearest neighbors that they share, to construct a shared nearest neighbor graph.33 Densely connected cells are defined as clusters by optimizing modularity with the Louvain community detection method.34
To predict the intercellular communication between ECs and functional cell types in each organ, we obtained ligand–receptor pairs compiled as previously described.35 We defined a ligand or receptor as “expressed” in a particular cell type if 25% of the cells of that type had at least 1 read count for the gene encoding the ligand/receptor. To define networks of cell-to-cell communication, we linked any 2 cell types if the ligand was expressed in the former cell type and the receptor in the latter. Lines connecting cell populations are colored according to the population broadcasting the ligand and are connected to the population expressing the receptor. Networks were plotted using the igraph R package.
Immunofluorescence was performed on 10-µm deparaffinized sections as previously described.36 Six C57Bl/6 mice 8 to 10 weeks of age (3 male and 3 female) were sacrificed under approved Stanford University administrative panel on laboratory animal care protocol 26923. All tissues were dissected and collected in phosphate-buffered saline, then fixed overnight in 4% paraformaldehyde at 4°C. The next day, tissues were washed in phosphate-buffered saline and PBT (phosphate-buffered saline containing 0.1% Tween-20), dehydrated in an ascending methanol sequence, xylene-treated, embedded in paraffin, and sectioned at 10 µm. Sections were subjected to antigen retrieval in Tris buffer pH 10.0 for 5 minutes, washed 3 times in 0.1% PBT, and incubated in blocking buffer (0.5% dried milk powder, 99.5% PBT) for 2 hours at room temperature. Primary antibodies were incubated in blocking buffer overnight at 4 °C with the following dilutions: Erg (Abcam, ab92513, 1:1000), Endomucin (Santa Cruz, sc-65495, 1:250), and Lars2 (leucyl-tRNA synthase 2; Proteintech, 17097-1-AP, 1:200). The next day, sections were washed 3 times with PBT and incubated for 1 hour with corresponding secondary antibodies at 1:500 dilution in blocking buffer at room temperature. After 3 washes in phosphate-buffered saline, DAPI (4′,6-diamidino-2-phenylindole; SigmaAldrich, 1:2000) was added to counterstain the nuclei. The sections were mounted using Prolong Gold Antifade Reagent (Invitrogen, P36934) and imaged using a Zeiss LSM 700 confocal microscope. All procedures were performed in accordance with the institutional guidelines of Stanford University Administrative Panel on Laboratory Animal Care.
DEGs among tissues and clusters were detected by comparing ECs in each tissue or cluster against all other ECs using a Wilcoxon rank sum test. A gene was defined as a cluster or tissue marker if it could be detected in ≥25% of cells, and the log-fold change in its expression was ≥2 between cells of cluster X and all other cells (adjusted P<0.05). For the analysis comparing male and female tissue-specific ECs, genes with an adjusted P<0.05 were deemed DEGs. We then defined the top DEGs as the 5 DEGs with the highest average log-fold change compared with ECs from the opposite sex. All these analyses were performed in the Seurat package v2.3. Kyoto Encyclopedia of Genes and Genomes pathway enrichment analyses were performed with geneAnswers R package.37
The Tabula Muris investigators performed scRNA-seq on 100 000+ cells in 20 major organs from 4 male and 3 female adult C57BL/6 mice.26 As described in the original article, cell type annotations were generated by identifying clusters within a given tissue that were defined by endothelial markers (Table I in the Data Supplement). Using these annotations, we extracted single EC transcriptome data in 12 organs (adipose tissue, aorta, brain, diaphragm, heart, kidney, liver, lung, mammary gland, pancreas, skeletal muscle, and trachea) with a sufficient number of ECs for downstream analyses (Figure 1A and 1D and Figure I in the Data Supplement). These cells have robust expression of endothelial genes, including Pecam1, Cdh5, Tie1, and Egfl7 (Figure IIA in the Data Supplement), with minimal expression of neuronal, kidney, and lung genes (Figure IIB, IIE, and IIF in the Data Supplement). Some ECs express hepatocyte (eg, Alb and Ttr; Figure IIC in the Data Supplement) and cardiomyocyte (eg, Tnnt2, Tnni3, and Nppa; Figure IID in the Data Supplement) markers, potentially reflecting the transdifferentiation potential of cardiac34,35 and hepatic sinusoidal ECs.38 In general, however, the majority of cells extracted from the Tabula Muris study show a uniformly strong expression of endothelial genes with little to no expression of parenchymal markers (Table II in the Data Supplement).
To determine whether scRNA-seq can capture tissue-specific EC transcriptomic signatures, we compared the Tabula Muris scRNA-seq EC data with a previously reported microarray dataset of ECs isolated from various mouse organs via intravital immunolabeling.20 We observed strong correlations between scRNA-seq and microarray-based tissue-specific EC gene expression (Figure IIIA–IIIF in the Data Supplement). Correlations of scRNA-seq and microarray-based gene expression measurements tend to be strongest when comparing ECs from the same tissue (Figure IIIG in the Data Supplement). We observed that when performing unsupervised hierarchical cluster analysis, tissue-specific ECs characterized by microarray clustered with Tabula Muris ECs from that same tissue (Figure IVG in the Data Supplement). The above analyses demonstrate that scRNA-seq can be used to effectively capture tissue-specific EC transcriptomic signatures from heterogeneous cell mixtures.
The Tabula Muris investigators observed 4 groups of transcriptomically distinct ECs when projecting cells from 20 mouse organs into 2-dimensional t-SNE space.26 These groups include liver, lung, and brain ECs as well as a heterogeneous mixture of ECs from various tissues. To better resolve differences between tissue-specific ECs, we visualized ECs alone in 2-dimensional t-SNE space (Figure 1B). Using this approach, we observed that ECs mostly segregate based on tissue of origin. ECs from some organs (eg, brain, kidney, lung, and liver) appear to have unique transcriptomic identities, whereas ECs from other organs (eg, adipose, heart, and aorta) show more overlap in gene expression as evidenced by uniform manifold approximation and projection (Figure 1B and 1C). These findings suggest that different organs may possess varying levels of functional specialization from ECs.
To define markers of tissue-specific ECs, we identified transcripts enriched in ECs from each of the 12 organs using the nonparametric Wilcoxon rank sum test. The top 10 DEGs in ECs from each organ are shown in Figure 2A, and a comprehensive list of DEGs is shown in Table III in the Data Supplement. Bolded DEGs encode cell surface proteins that may be useful for targeted delivery of therapeutics to specific tissues. The most specialized DEG profiles were found in brain and pancreas ECs, which expressed mainly solute carrier transporters and digestive enzymes, respectively. To visualize the tissue specificity of DEGs, we generated a heatmap displaying the expression of tissue-specific DEGs across all 12 organs. As seen in the t-SNE presentation (Figure 1B), brain and liver ECs express unique sets of genes that are virtually undetected in ECs from other organs (Figure 2B). Conversely, the DEGs characterizing heart, diaphragm, adipose tissue, skeletal muscle, and mammary gland ECs are less specific (Figure 2B). We also found that, with the exception of liver ECs, the expression of tissue-specific EC DEGs identified previously by microarray20 correlated well between microarray and scRNA-seq–based transcriptomic measurements (Figure IVA–IVF in the Data Supplement). These findings validate the utility of scRNA-seq for identifying the transcriptomic signature of tissue-specific ECs.
Using the identified organ-specific EC DEGs, we determined molecular pathways unique to organ-specific ECs by performing gene set enrichment analysis with the Kyoto Encyclopedia of Genes and Genomes pathway database. Pathways associated with the greatest number of DEGs in ECs were visualized in chord plots for each of the 12 organs (Figure 3A). Gene enrichment analysis identified both commonly shared and unique molecular and cellular pathways in ECs from different organs. Examples of unique pathways include osteoclast differentiation in adipose tissue ECs, ErbB signaling in brain ECs, axon guidance in cardiac ECs, and endocytosis in kidney ECs. Major developmental pathways known to play a critical role in endothelial homeostasis and function such as the Wnt, MAPK (mitogen-activated protein kinase), cytokine–cytokine receptor interaction, and metabolism-associated pathways were commonly found in multiple organs. The majority of the genes associated with the common pathways in organ-specific ECs were unique (Table IV and Figure V in the Data Supplement). For instance, Wnt signaling was found to be highly upregulated in ECs from the brain, heart, liver, and lung, but the genes in organ-specific ECs contributing to Wnt signaling regulation were different. Namely, brain ECs show a higher expression of Axin2, Fzd6, and Nkd1, whereas cardiac ECs expressed Ccnd1, Ctnnbip1, and Plcb4, and liver ECs expressed Apc, Ep300, and Lrp6. Likewise, we identified unique DEGs specific to ECs in each organ that are associated with MAPK signaling, cytokine–cytokine receptor interaction, and metabolic pathways.
scRNA-seq of all cells collected from a given tissue enables gene expression comparison and analysis among individual cell types. A major known role of ECs in tissue homeostasis and function is mediated through the secretion of EC-specific paracrine factors called angiocrine factors.39 We sought to unveil unique angiocrine factors expressed by ECs in each organ. From the EC single-cell transcriptome, we identified potential ligand–receptor pairs between ECs and all sequenced parenchymal cell types in each of the 12 organs as previously described35,40 (Figure 3B, left). A comprehensive list of the identified ligand–receptor pairs is provided in Table V in the Data Supplement. This analysis revealed the existence of unique angiocrine ligand–receptor pairings between ECs and parenchymal cells in each organ. For example, Efna1 expressed in brain ECs is known to bind to Epha5 and Pehb5 gene-coding receptors in neurons and oligodendrocyte precursor cells, whereas Edn3 expression from ECs is projected to interact with Ednra in brain pericytes. Similarly, we depicted unique angiocrine and paracrine relationships in 8 major organs (Figure 3B, right).
We used unsupervised clustering to identify novel subpopulations of ECs that may be independent of the tissue of origin. Specifically, we found 13 unique clusters using a graph-based clustering approach (Figure 4A and 4B). We then determined the percentage (Figure 4C) and the absolute numbers (Figure VIA in the Data Supplement) of ECs from each organ that comprised the 13 clusters, with the enriched gene list for each cluster provided in Table VI in the Data Supplement. Certain clusters show enrichment of ECs from a single organ, whereas other clusters are composed of ECs from various organs (Figure 4C and Figure VIA in the Data Supplement). We analyzed the proportion (Figure 4D) and the absolute numbers (Figure VIB in the Data Supplement) of ECs in each organ that were assigned to the various clusters. Pathway enrichment analysis and cluster-specific DEGs allowed us to infer the biological identity of the unsupervised clusters (Figure 4E through 4G and Figure VII in the Data Supplement). For example, cluster 4 is a subpopulation of lung ECs that overexpresses transcripts involved in antigen processing and presentation, allograft rejection, and graft-versus-host disease (Figure 4E), leading us to hypothesize this subpopulation may represent antigen-presenting ECs. Cluster 5 comprises ECs from a number of different organs that show enrichment of transcripts involved in toxoplasmosis, cytosolic DNA sensing, and MAPK signaling and therefore may be ECs responding to an infection (Figure 4F). Cells in cluster 10 express genes in the stromal or smooth muscle cell lineage (Table VI and Figure VII in the Data Supplement) and may represent ECs undergoing endothelial-to-mesenchymal transition. We found that ECs in cluster 9, a highly unique cluster in the transcriptome, had high expression of lymphatic markers such as Prox1, Pdpn, Lyve1, and Flt4 (Figure 4G and Figure VIC in the Data Supplement). We combined scRNA-seq with unsupervised clustering to identify EC subtypes that either reside within a single organ or in multiple organs.
Because the Tabula Muris investigators analyzed both male and female mice, we probed for sex-dependent EC gene expression in brain, heart, lung, adipose tissue, aorta, and kidney (Figure 5A and Figure VIIIA through VIIIE in the Data Supplement). To ascertain the influence of sex on EC gene expression, we performed unsupervised clustering to determine whether ECs cluster by sex within a given organ (Figure IX and Table VII in the Data Supplement). Some tissues (eg, aorta, brain, lung) had subclusters of ECs that were composed almost entirely of ECs from a single sex whereas other tissues (adipose tissue, heart, kidney) had a more even sex distribution between subclusters. These findings suggest that there are subsets of ECs within certain tissues that are sex-specific. We also determined the top DEGs in male versus female organ-specific ECs (Figure 5B and Figure VIIIF through VIIIH in the Data Supplement) and used them to perform pathway enrichment analysis (Figure 5C). Analysis of the overlap between DEGs identified in male versus female tissue-specific ECs indicated that the sex dependency of DEGs varies among different tissues (Figure X in the Data Supplement). Notably, we found Lars2, a non–sex-linked gene coding for leucyl-tRNA synthase 2, to be robustly and uniformly expressed in all male ECs, whereas its expression is low and highly variable in female ECs (Figure 6A and 6B). Whereas Lars2 gene expression is relatively ubiquitous among various cell types (data not shown), LARS2 protein expression appeared largely specific to ECs in most organs (Figure 6C). Similar to mRNA, LARS2 protein expression was enriched in male ECs, especially in the brain, heart ventricle, lung, and liver (Figure 6C). Taken together, these findings indicate that sex is an important source of transcriptomic variation in tissue-specific ECs.
To determine the relevance of our findings in murine ECs to human endothelial biology, we first compared the Tabula Muris dataset with bulk RNA sequencing measurements made in ECs isolated from human fetal organs (eg, heart, kidney, liver, and lung).23 We found that the correlation of adult murine and fetal human EC gene expression tends to be strongest when comparing ECs from the same organ (Figure 7A). Furthermore, markers of tissue-specific ECs identified in mice were also enriched in their corresponding human tissue-specific ECs (Figure 7B). These findings indicate that the transcriptomic manifestation of tissue specificity in ECs is relatively well conserved between mice and humans. We checked the expression of the top 10 murine tissue-specific EC DEGs in human induced pluripotent stem cell– and human embryonic stem cell–derived ECs and their progenitors generated previously by our group.40,41 Progenitor induced pluripotent stem cell ECs showed an elevated expression of genes enriched in the aorta, adipose tissue, heart, and liver ECs (Figure 7C, middle panel). Fully differentiated induced pluripotent stem cell ECs and embryonic stem cell ECs were most notably characterized by the expression of aorta, heart, liver, and mammary gland EC genes, whereas lacking expression of genes in the brain and pancreas ECs (Figure 7C, right panel). Thus, the human pluripotent stem cell–derived ECs do not mimic ECs from any one particular tissue with respect to global gene expression, necessitating further development of methodologies to generate induced pluripotent stem cell ECs or embryonic stem cell ECs that possess organ-specific transcriptomic profiles to enable the most effective use in disease modeling or cell-based therapy applications.
The vasculature is present in all major organs, underpinning homeostasis and function throughout the body. It is versatile in its role to accommodate to the unique physiologic function of each organ, such as in nutrient transport, endocrine signaling, waste disposal, and disease protection.7 However, genes and molecular pathways that govern the organ-specific role of ECs have not been clearly defined, primarily attributable to insufficient methodologies to investigate a single cell type from various tissues in parallel. Previous attempts to decipher functional and transcriptomic features of organ-specific ECs were insightful but not comprehensive, suffering from a limited number of organs analyzed in parallel or from bulk analysis masking detection of small populations or lowly expressed genes.18,20
Recent advances in scRNA-seq have resolved these limitations, aided by the generation of transcriptomic and epigenetic atlases of major mammalian organs. Kalucka and colleagues10 published a comprehensive atlas of single-cell transcriptome measurements made in murine ECs from various tissues. With respect to tissue-specific EC markers, the overlap between studies was high for the brain, liver, and kidney, medium for the lung, and low for the heart. This finding highlights the necessity for comparing multiple studies to identify robust markers of tissue-specific ECs. These genes, especially those that code for membrane-surface proteins, are promising targets for organ-specific delivery of small-molecule chemicals and DNA- or RNA-based therapeutics that possess greater target specificity with reduced risk of side effects.
Because the Tabula Muris consortium used both male and female mice, we were able to look for sex differences in tissue-specific EC gene expression. We found that markers of tissue-specific ECs, enriched pathways, and endothelial subpopulations varied between male and female mice, especially in the brain, heart, and lung. This transcriptomic variation may, in part, drive known sex differences in endothelial biology42 and cardiovascular disease risk.43 Furthermore, when designing tissue-specific drug delivery strategies, it will be vital to ensure that EC membrane targets are robustly expressed in both males and females. We also discovered that Lars2 gene coding for mitochondrial leucyl-tRNA synthetase was highly enriched in male compared with female ECs. Interestingly, LAR2 variants are associated with multiorgan dysfunction with varying phenotypes between males and females.44 Sex differences in endothelial Lars2 expression may contribute to this phenotypic variation.
We also compared global gene expression between adult mouse and fetal human tissue-specific ECs as determined by single-cell and bulk RNA sequencing, respectively. We found that, despite differences in the developmental stage, there was considerable overlap in gene expression between ECs from the same tissue. This suggests that, to a certain extent, the transcriptional manifestation of tissue specificity in ECs is established during development. Multi-institutional collaborative efforts are underway to establish human cell atlases, including the Human Cell Atlas45 and the Human BioMolecular Atlas Program,46 whose data when generated can similarly be analyzed to decipher tissue-specific transcriptomic features of human ECs. Specifically, the direct juxtaposition of mouse and human EC single-cell gene expression profiles will be critical in identifying similarities and differences between tissue-specific ECs attributable to species, age, and sex. As there are differences between mice and humans with respect to vessel organization, morphology, and hemodynamics, we expect that this variation will be reflected at the transcriptomic level. The completed human scRNA-seq atlas will be valuable in addressing these questions.
Future studies will delve into tissue-specific differences in arterial, venous, and lymphatic specification of blood vessels and identify transcriptomic and functional differences in larger vessels versus capillary ECs. The origin of vascular ECs has long been a topic of debate, with a plethora of lineage tracing models and cell sorting performed to decipher the identities of stem/progenitor cell populations that give rise to the endothelium. Moreover, adult ECs innately possess a high level of plasticity, capable of undergoing endothelial-to-mesenchymal transition in various tissue types and disease states.47 Further analysis of scRNA-seq datasets of ECs from all organs during embryonic development or in association with aging and disease conditions will provide profound and novel information in addressing these questions.48
Supported by the National Institutes of Health grant K99 HL150216 and Stanford Cardiovascular Institute Seed Grant (Dr Paik); American Heart Association Postdoctoral Fellowship 20POST35210924 (Dr Tian); National Institutes of Health grant T32 HL098049 (Dr Williams); American Heart Association Undergraduate Summer Research Program (R. Mishra); and National Institutes of Health grants R01 HL145676, R01 HL113006, R01 HL123968, and R01 HL141851, Tobacco-Related Disease Research Program 27IR-0012, and American Heart Association grant 17MERIT33610009 (Dr J.C. Wu).
Dr Wu is a cofounder of Khloris Biosciences but has no competing interests, as the work presented here is independent. The other authors report no conflicts.
Data Supplement Tables I–VII
Data Supplement Figures I–X
*Drs Paik, Tian, and Williams contributed equally.
Sources of Funding, see page 1861
The Data Supplement is available with this article at https://www.ahajournals.org/doi/suppl/10.1161/circulationaha.119.041433.
联系客服