A novel lesional transcriptional signature separates atherosclerosis with and without diabetes in Yorkshire swine and humans

Abstract

Objective:

Accelerated atherosclerosis in diabetes mellitus constitutes an ongoing challenge despite optimal medical therapies. This study aimed to identify evolutionarily conserved lesion-based regulatory signaling networks in diabetes versus non-diabetes conditions during the development of atherosclerosis in an initial translational effort to provide insights for targets.

Approach and Results:

Serial 3-mm coronary artery segments of hypercholesterolemic Yorkshire swine and diabetic-hypercholesterolemic swine were characterized as mild, moderate, or severe phenotypic manifestations of coronary atherosclerosis based on histopathologic examination. Lesional RNA-Sequencing was performed (n=3-8 lesions per group) corresponding to increasing phenotypic severity. Differentially expressed genes, transcription factors, upstream regulators, and hubs were validated using NanoString technology and a human atherosclerotic specimen cohort.

Despite similar stage histopathological characterization of lesions, genome-wide transcriptomics revealed gene sets and nodal signaling pathways uniquely expressed in diabetic lesions including signaling pathways for Th17, IL-17F, TWEAK, CD27, and PI3K/Akt. In contrast, pathways of non-diabetic lesions involved TREM-1 and Th1 and Th2 responses during the initiation stage, whereas networks for mitochondrial dysfunction, oxidative phosphorylation, and lipid metabolism emerged with progression. RNA-Seq data was validated in a human atherosclerosis specimen cohort using machine learning algorithms. F8, MAPKAPK3, and ITGB1 emerged as powerful genes for clustering diabetic vs. non-diabetic lesions and for separating different degrees of atherosclerosis progression.

Conclusions:

This study identifies evolutionarily conserved gene signatures and signaling pathways in a stage-specific manner that successfully distinguishes diabetes and non-diabetes-associated atherosclerosis. These findings establish new molecular insights and therapeutic opportunities to address accelerated atherosclerotic lesion formation in diabetes.

Graphical Abstract

INTRODUCTION

Cardiovascular diseases (CVD) comprise the most frequent cause of death representing approximately 30% of all deaths worldwide.¹ The most prevalent complications of atherosclerotic lesions of medium to large-sized arteries are myocardial infarction and ischemic stroke.² One major driver for this trend is the ongoing epidemic of obesity-induced insulin resistance and type 2 diabetes.³ Moreover, insulin resistance represents an independent risk factor for CVD.⁴ Despite the vast clinical experience linking diabetes to vascular disease, mechanistic gaps remain connecting hypercholesteremia and hyperglycemia to the accelerated atherosclerosis observed in subjects with diabetes.⁵ Therefore, identification of evolutionarily conserved lesion-based regulatory signaling networks in diabetes versus non-diabetes conditions may offer a better understanding of the accelerated progression of atherosclerosis with diabetes.

While genetically modified mice (e.g. Ldlr^−/− or ApoE^−/− mice) have served as powerful tools to study the role of any given gene, mice are limited by their exaggerated hypercholesterolemia lack of spontaneous coronary artery disease, rare development of plaque disruption, and characteristic differences in lesion composition such as the lack of a thick fibrous cap seen in chronic human atherosclerosis.^6–9 Such differences between rodents and humans have made it incumbent for additional tools to be used, such as pig models. Furthermore, most human lesion-based studies are not derived from coronary arteries, but larger ones (e.g. aorta or carotid arteries) due to availability, thereby creating a void for coronary-specific targets and molecular insights. This point has importance given the distinct embryological origins of the coronary arteries from other vessels. ¹⁰ Pigs are not just phylogenetically closer to humans, but lesions also resemble human coronary lesions in composition and complexity with a thick fibrous cap, calcification, and features of microhemorrhage.¹¹ Moreover, pigs rendered diabetic by streptozotocin (STZ) administration in combination with a high cholesterol diet (HCD) exhibit complex lesions in the first few centimeters of the coronary arteries without a major effect on plasma cholesterol,¹² thereby providing a translational opportunity to explore diabetes-associated coronary atherosclerosis in larger animals.

This study explored the hypothesis that lesional differences in the transcriptomic profile are more pronounced compared to histopathological markers in response to systemic risk factors as hypercholesterolemia and diabetes. Accounting for the complex interplay of genes that may operate in a stage-specific manner may yield a more complete understanding of processes that control initiation and progression of atherosclerosis under non-diabetic and diabetic conditions. To address this possibility, we compared transcriptomic profiles of coronary plaques in non-diabetic and diabetic Yorkshire swine, which were characterized and separated in groups based on their stage-specific burden using established histopathological markers. These morphologic features allowed us to perform lesion-based RNA-Seq analyses from RNA-derived from mild, moderate and severe atherosclerotic lesions in non-diabetic pigs and from moderate and severe lesions from diabetic-rendered pigs. This discovery study identified differentially expressed genes, transcription factors, upstream regulators and hubs that contribute to specific signaling networks in diabetic and non-diabetic lesion progression. A defined selection of those genes was validated by transcriptomic NanoString profiling in human atherosclerotic specimens and prospectively classified lesions as diabetic or non-diabetic based on their gene profile.

METHODS

Data Availability

Anonymized data and materials have been made publicly available and can be accessed at GSE162391.

Animals

All protocols concerning animal use were approved by the Institutional Animal Care and Use Committee at Brigham and Women’s Hospital and Harvard Medical School, Boston, MA and conducted in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals. The study protocol included 5 male hypercholesterolemic Yorkshire swine and 2 male hypercholesterolemic swine (8 to 12 weeks old) who were also rendered diabetic by administration of i.v. STZ and followed over 60 weeks as described.¹³ Male swine were used in this study to recapitulate the hypercholesterolemic and diabetic conditions achieved as previously described¹². Detailed sectioning of 3-mm coronary artery segments from the left anterior descending, left circumflex, and right coronary artery was performed so that the gene sequencing samples were derived from the exact same portions of the coronary artery plaques used for the histology and IHC analyses.

Human Specimen Cohort

Frozen sections were prepared from human normal carotid arteries and carotid atherosclerotic lesions that were obtained from the Division of Cardiovascular Medicine, Brigham and Women’s Hospital in accordance with the Institutional Review Board-approved protocol for use of discarded human tissues (protocol #2010-P-001930/2).

NanoString

Customized nanoString for 55 genes was designed based on the pig genome with possible conservation to human genome. NanoString assay was performed based on manufacturer’s protocol by Molecular Genetics Core Facility at Boston Children’s Hospital. Pig samples were normalized to negative control genes as suggest by manufacturer, whereas human data was normalized with edgeR.

Immunohistochemistry

Histology analyses included H & E, van Gieson elastin staining, smooth muscle cell α-actin, oil red-O staining (ORO), picrosirius red staining and CD45 cells.¹⁴ Detailed information for all the antibodies used in this study can be found in the Major Resources Table. Based on histopathologic characteristics (i.e. intima-to-media (IM) ratio, severity of lipid accumulation, and inflammation), the lesions from pig tissue samples were classified into three categories representing distinct stages of the natural history of atherosclerosis as proposed by Virmani et al¹⁵: (a) mild lesions characterized by minimal depositions of lipids and inflammatory cells into the intima (Intimal Thickening by Virmani), (b) moderate lesions characterized by larger masses of lipid-laden inflammatory cells without evidence of fibrous cap and IM≥0.15 (Pathologic Intimal Thickening by Virmani), and (c) severe lesions with a severely inflamed fibrous cap often overlying a necrotic lipid core (Fibrous Cap Atheromata and Thin Fibrous Cap Atheromata by Virmani). Lesions characterized as mild, moderate, and severe histopathologically were derived from each of the animals, thereby highlighting that this study was lesion-centric and not animal-specific analyses.

RNA Extraction and Quality Control

RNA was isolated from 3-mm coronary artery cross-sections using TRIzol Reagent based on manufacturer’s protocol (Ambion). Quality control of RNA was examined using a combination of Bioanalyzer (Agilent) and Qubit (ThermoFisher).

Canonical Pathway Analysis and Upstream Regulator

For canonical pathway analysis and the prediction of upstream regulators Ingenuity Pathway Analysis (IPA) software (Qiagen) was used. The pathway activity (z score) was computed to determine whether the activity of canonical pathways is increased or decreased on the basis of differentially expressed genes in the data sets. The significant values for the canonical pathways were calculated by Fisher’s exact test right-tailed. Pathways with a p-value <0.05 were included for the analysis of hubs and radar graphs.

Prediction of biological function of canonical pathways

BioFun package in R program was used to search for the involvement of each IPA canonical pathway in the Biological Function Classification Database of IPA known as “Ingenuity canonical pathway” and counts the number of pathways involved in a specific biological function. The results are illustrated as radar graphs as described in.¹⁶

Hierarchical Clustering and Heatmaps

The function heatmap2 in the R package gplots or GENE-E R package were used to generate the hierarchical clustering and the associated heat maps for mRNA sequencing and NanoString data. Based on Pearson correlation method, pairwise correlation matrix between items was computed and converted to a distance matrix; ultimately, clustering was computed on the resulting distance matrix. Average linkage method used average to calculate the distance matrix. To draw simple Venn diagrams without size adjustments, Venny (Oliveros, J.C. (2007-2015) Venny. An interactive tool for comparing lists with Venn’s diagrams. https://bioinfogp.cnb.csic.es/tools/venny/index.html and jvenn were used.¹⁷

Gene Clouds (Word Clouds)

To summarize and picture large amounts of gene enrichment data from a pathway analysis data set and discover biological patterns, a cloud was produced with Wordle.net and Word cloud R package. The font size of a gene (tag) is determined by its incidence in the pathway analysis data set.

Volcano plots

Volcano plots were created in R environment using EnhancedVolcano package (Blighe K, Rana S, Lewis M (2019). EnhancedVolcano: Publication-ready volcano plots with enhanced coloring and labeling. R package version 1.4.0, https://github.com/kevinblighe/EnhancedVolcano.).

Principal component analysis (PCA)

For PCA, log2 fold change of mRNAs were used as an input for prcomp function in R (R Core Team, 2016) and FactoMineR package. rgl, and scatterplot3d R package were used for the PCA of 3D plots. To attain better numerical accuracy, the computation was performed utilizing a singular value decomposition of the (centered and scaled) data matrix.

3D ellipsoid chart and point identification for biomarkers

The scatter3d() function in car (Companion to Applied Regression) package was used in order to call for the rgl package, to draw 3D scatter plots with various regression surfaces (http://cran.r-project.org) using XQuartz (The X Window System, version 2.7.9). The display of the surfaces was fitted using a linear method.

Circos plots

All the circus plots were created using circos-0.69-9 software on UNIX environment.¹⁸ To ensure that cell values are read in and processed, percentile cutoff was adjusted to zero.

Correlation plots

Corrplot package (ver 0.84) in R was used to generate the visualization of correlation between sequencing and NanoString data (Taiyun Wei and Viliam Simko (2017). R package “corrplot”: Visualization of a Correlation Matrix (Version 0.84). Available from https://github.com/taiyun/corrplot.

RNA-Seq DESeq2

RNA-Seq analysis was performed after ribodepletion and standard library construction using Illumina HiSeq2500 V4 2x100 PE (Genewiz, South Plainfield, NJ). After low abundant quality bases at ends were trimmed, sequences were mapped to Sus Scrofa reference genome (V.10.2) using CLC Genomics workbench. Differential expression analysis between specified groups was performed using DESeq2.¹⁹

Statistical Analysis

For all analyses, normality and equal variance was tested to determine whether the applied parametric tests were appropriate. For all correlation analysis normality of data was tested using Shapiro-Wilk test and qq-plots method. As all data sets for correlation were normally distributed, Pearson correlation coefficient was applied. Calculations of significance for IHC images used one-way ANOVA with post hoc Tukey’s test for correction of multiple comparisons (ns, not significant; *P<0.05; **P<0.01; ***P<0.001; ****P<0.0001).

RESULTS

Diabetic lesions show no significant changes for histopathological markers compared to non-diabetic lesions.

Similar to humans, cholesterol-rich diet in pigs induces multifocal atherosclerotic lesions with a similar composition of histological features including development of inflammatory cell infiltrate, necrotic core, and thick fibrous cap.^20,21 To explore a lesion-based histopathological characterization, Yorkshire swine consumed a high cholesterol diet (HCD) for 60 weeks (n=5 swine) and a second group was rendered diabetic by i.v. injection of the β-cell cytotoxin streptozotocin (STZ) in combination with HCD for 60 weeks (n=2 swine) (Figure 1A). Each animal developed hypercholesterolemia, with serial cholesterol values in the range of 400-500 mg/dL (monitored every 4 weeks; Suppl. Figure IA). STZ injection leads to the development of full metabolic manifestations of diabetes with serum glucose values >350mg/dL (Suppl. Figure IB) consistent with previous studies.^12,22,23 The average triglycerides were 73.2 mg/dl in the DM swine, and 54.2 mg/dl in the non-DM swine. Sequential 3-mm coronary artery segments of non-diabetics (n=13 lesions from 5 swine) and diabetics (n=10 lesions from 2 swine) were characterized based on lesional histopathological markers such as Oil Red O (ORO), intima/media-ratio (I/M-Ratio), plaque-IEL, CD45, elastin and smooth muscle cell (SMC) composition in mild, moderate and severe atherosclerotic plaques (Figure 1B,C Suppl. Table I). As expected, these markers increased with progression of atherosclerosis and, for ORO and CD45 expression, positively correlated across the samples (Figure 1D). In support of accelerated atherosclerosis with diabetes, the coronary arteries of diabetic pigs contained ~2.5-fold greater proportion of plaques (average 0.458 plaques per segment per animal) compared to the non-diabetic pigs (average 0.185 plaques per segment per animal). Moreover, while there were mild, moderate, and severe lesions detected in the non-diabetic group, the diabetic group contained only moderate to severe lesions.

Figure 1. Histopathological characterization of atherosclerotic lesions in non-diabetic and diabetic swine.

(A) Two groups of Yorkshire swine (n=2-5 animals) were placed on a high cholesterol diet (HCD) for up to 60 weeks, where one group was rendered diabetic following injection of streptozotocin. Coronary artery segments were characterized for their severity of disease progression based on histopathological markers such as (B) Oil Red O (ORO) staining, intima-media (I-M) ratio, plaque-internal elastic lamina (IEL) and (C) CD45, SMC, and elastin. (D) Pearson correlation of markers for ORO and CD45% (r²=0.623, p=0.009). (one-way ANOVA; ns, not significant; *P<0.05; **P<0.01; ***P<0.001; ****P<0.0001).

Distinct genome-wide transcriptomic profile of non-diabetic vs. diabetic atherosclerotic lesions.

For the purpose of this lesion-based study, adjacent coronary artery segments were selected for RNA-Seq. Unsupervised hierarchical cluster analysis showed distinct separation between non-diabetic and diabetic atherosclerotic lesions following geometric DESeq2 normalization (Figure 2A). Principal component analysis (PCA) of the whole transcriptome confirmed this distinct clustering of diabetic to non-diabetic lesions (Figure 2B). As established in the initial study design, the comparisons (Figure 2C) were explored to identify differentially expressed genes (Reads>100, log2FC(0.5), FDR<0.1) across specific diabetic and non-diabetic lesion-based groups: initiation of atherosclerosis in non-diabetics (−/− DM; moderate vs. mild); progression of atherosclerosis in non-diabetics (−/− DM; severe vs. mild); moderate atherosclerosis in diabetics vs non-diabetics (+/− DM; moderate DM vs. moderate); and severe atherosclerosis in diabetics vs non-diabetics (+/−DM; severe DM vs. severe) (Figure 2C, Suppl. Table I).

Figure 2: Transcriptomic profiling of coronary artery diabetic and non-diabetic lesions.

(A) RNA derived from adjacent histopathological characterized coronary artery segments was used for genome-wide RNA-Seq analysis. Unsupervised hierarchical heatmap cluster analysis of all lesions separates mild, non-diabetic, and diabetic lesions. (B) Principal component analysis (PCA) of the whole transcriptome shows distinct clustering of non-diabetic and diabetic lesions. (C) Overview and terminology for different comparisons of data sets. (D) Heatmap of top five induced and reduced genes for each comparison. Volcano plots displaying differentially expressed genes for (E) non-diabetic progression −/− (comparing severe vs mild non-diabetic groups); and (F) diabetic moderate +/− (comparing moderate lesions in diabetic and non-diabetic groups). (G) Venn diagram for differentially expressed genes across the four comparisons. Top 10 up and down regulated genes common in (H) non-diabetic and (I) diabetic lesions.

Among the top regulated genes were mainly long non-coding RNA (lncRNA) transcripts and to a lesser extent protein coding genes such as S100A12, MMP9, MYH6, CHI3L1, and MMRN2 (Figure 2D). All significantly expressed genes are displayed in volcano plots (threshold log2FC(0.5), FDR<0.1) (Figure 2E,F, Suppl. Figure IIIA,B). The intragroup heterogeneity of severe vs. moderate atherosclerosis (i.e. advanced −/− and advanced +/+) did not yield any major changes in transcriptomics profile (Suppl. Figure III). Specific differentially expressed genes were identified by overlaying the comparisons (Figure 2G). Top ten up- and down-regulated genes only in initiation and progression non-diabetic lesions (n=347) such as ATPV0D2, ACP5, MMP12, NEB and DAPL1 are just a few among several other coding and non-coding transcripts (Figure 2H). Diabetic lesions showed overall more significantly expressed genes such as LOC100624650, AKAP6, LOC102167708 and PIK3R5. Genes with highest fold change in diabetic lesions were IL-8, FOS, CCL8, CKM, MB1, and TNNT2 (Figure 2I). Taken together, these initial transcriptomic comparisons revealed unique gene expression profiles of diabetic versus non-diabetic lesions despite few if any differences observed by the histopathological markers tested (Figure 1, Suppl. Figure II).

Canonical pathway analysis identifies mitochondrial dysfunction and PI3K signaling among top networks.

Projections that diabetes will increase within the next two decades by over 70% to afflict 333 million worldwide render urgent the quest to identify the regulatory signaling networks specific to diabetes lesions.²⁴ Extensive research has highlighted the importance of inflammation, remodeling of extracellular matrix, and lipid deposition in the arterial wall as some of the key processes in the development of atherosclerosis;^25,26 however, we have lacked a comprehensive lesion-based analyses in the presence and absence of diabetes.

Transcriptomics comparison of moderate non-diabetic lesions to mild non-diabetic lesions showed enriched signaling pathways in TREM-1, Th1 and Th2 signaling, and mediators of leukocyte trafficking (Figure 3A). With progression of atherosclerosis in non-hyperglycemic pigs, mitochondrial dysfunction, oxidative phosphorylation, sirtuin signaling, TCA Cycle II, phagosome maturation, and phagosome formation were among the most regulated networks (Figure 3B). When comparing diabetic moderate and severe lesions to non-diabetic moderate and severe lesions, signaling for ILK, IL-10, STAT3, CD40, NF-κB, sphingosine-1-phosphate, Tec kinase, and Rho family GTPases emerged. In addition, disease-specific signaling networks for hepatic fibrosis, rheumatoid arthritis, and osteoarthritis emerged under diabetic conditions (Figure 3C, D).

Figure 3. Canonical pathway analysis and their biological hubs for the different data sets.

(A-D) For each dataset, independent canonical pathway analysis was performed to identify top enriched signaling pathways with a predicted ‘activation’ or ‘inhibition’ or ‘no directionality’. Gene clouds reflect most frequent genes (i.e. hubs) that are most common in a given dataset across the significantly regulated canonical pathways (E). (F) Identification of stage-specific hubs with their expression across all five groups.

The gene hubs that drive the identified canonical pathways for each data set might be repetitive (i.e. present in more than one pathway). For example, among 127 pathways in the initiation−/− dataset in non-diabetic lesions, a total of 1173 repetitive hubs are involved, of which 229 hubs are unique with the highest frequency shown in the word cloud (Figure 3E). As such, we identified for each non-diabetic and diabetic stage of disease progression an exclusive list of hubs that we summarized in a heatmap (Figure 3F, Suppl. Table II). Predominantly present across all diabetic and non-diabetic lesions were hubs associated with PI3K signaling such as PIK3R4, PIK3CD, PIK3C2G, PI3KR5, CXCR4, LIMK1, and CD40, whereas hubs such as NFKB1, MAP2K2, MAPK10, NFKBIB, and FGFR4 were specific for diabetic lesions (Figure 3F).

Collectively, these unbiased canonical pathway analyses revealed at early stages of lesion development mainly inflammatory-associated signaling pathways, whereas with lesion progression mitochondrial dysfunction and phagosome maturation dominated. Diabetic lesions exhibited more progressive disease-linked pathways with enrichment of new inflammatory signaling pathways such as STAT3, CD40, NF-κB, S1P, Tec kinase, and Rho GTPases, compared to non-diabetic lesions. In contrast, hubs for PI3K signaling emerged as the top genes across signaling pathways common to all non-diabetic and diabetic lesions.

Broad PI3K activation and specific Th17 response under diabetic atherosclerosis lesion progression.

To integrate better stage-specific signaling networks based upon their directionality for activation or inhibition, we performed comparative analyses of signaling networks in non-diabetic versus diabetic lesions using their z-scores. Across all comparisons, signaling pathways inhibited included PPAR, PPARα/RXRα activation, apelin cardiomyocyte signaling, and inhibition of matrix metalloproteases, whereas NFκB, interleukin-8 (IL-8), TREM-1, thrombin signaling, and PI3K downstream pathways such as Rac signaling, NFAT, CD28 in T helper cells, and PI3K in B lymphocytes were all activated (Figure 4A). Specific to non-diabetic lesions were signaling pathways associated with oxidative phosphorylation, sirtuin signaling, protein kinase A, and RhoA signaling. Diabetic lesions displayed a particular signature for pathways involved in TNF related weak inducer of apoptosis (TWEAK), CD27, and activation in Th17 pathway and IL-17F (Figure 4A). To place the identified disease- and stage-specific pathways in a broader biological context, radar graphs were created to summarize all significantly regulated canonical pathways for their biological functions and cellular processes involved (Figure 4B, Suppl. Figure IV). Interestingly, the biological functions of ‘Cancer’ and ‘Apoptosis’ were more enriched under diabetic conditions specifically for bladder, colorectal, and breast cancer, which was driven by increased activation of ERK/MAPK, JAK/STAT, IL15, and Her2 signaling. The cellular ‘Growth Proliferation and Development’ function was more pronounced under diabetic conditions due to increased ILK, PEDF, Rank, Cdc42, FGF, AMPK, and increased epithelial tight junction signaling. In addition, ‘Intracellular Signaling’ and ‘Metabolic Pathways’ was enriched in diabetic lesions as a result of increased Erk5, RhoA, Pak, and Calcium signaling (Figure 4B, Suppl Figure IV). Consistent with previous signaling analysis (Figure 3), sirtuin signaling, a pathway involved in DNA repair and aging, was enriched in non-diabetic lesions but was markedly reduced in diabetic lesions. A PI3K signature appeared again across all non-diabetic and diabetic groups. Diabetic lesions showed a specific signaling network for Th17 and IL-17F as well as TWEAK signaling. Taken together, diabetic lesions exhibited a “cancer-like” gene signature distinct from non-diabetic lesions.

Figure 4. Comparative signaling pathway analysis.

(A) Ingenuity Pathway Analyses (IPA) identifies common and different signaling pathways between increasing severity of lesions with (+/−) or without (−/−) diabetes. (B) Canonical pathways are categorized in broader biological functions and the number of signaling pathways for each function is shown in a radar graph comparing progression in non-diabetics (−/−) versus severe lesions in diabetics (+/−).

Identification of transcriptional upstream regulators at lesion-specific stages.

The complex transcriptomic changes that affect signaling pathways and accompany atherogenesis in non-diabetic and diabetic plaque formation can be controlled in a variety of ways. To explain changes on common and unique gene hubs, we explored whether there are common upstream regulators at lesion-specific stages in non-diabetes and diabetes conditions. IPA predicts upstream regulators from genes that directly affect the expression of other genes as well as indirect nodes (Figure 5A). The direction of regulation can be either “activating” or “inhibiting” based on literature-derived findings, which is indicated by a z-score. In our analysis, stage-specific upstream regulators were only included when their prediction overlaps with actual fold changes in our data sets (Figure 3B). In early onset of non-diabetic atherosclerosis, VDR and SPP1 were the only two predicted upstream regulators (Figure 5C). Upstream regulators described in lipid metabolism such as CYPSI, OLR1, APOE, and FDFT1 emerged with progression of non-diabetic atherosclerosis. Additionally, upstream regulators described for signaling networks involved in mitochondrial dysfunction and oxidative phosphorylation include FOXO1, NSUN3, GPX1, and NCF1 under non-diabetic lesion progression (Figure 5C–E). In contrast, in the diabetic lesions reflecting progression of atherosclerosis, CXCR4, STAT3, IL-8, IL1-A, CD38, KLF4, PTGS2, and IGF1 showed distinct expression compared to non-diabetic lesions (Figures 5F–H).

Figure 5. Stage-specific identification of upstream regulators and transcription factors.

(A) Changes of gene signatures and signaling pathways may be controlled by upstream regulators and/or transcription factors. (B) IPA for upstream regulators was filtered by those hits that were differentially expressed in our RNA-Seq data sets. (C-H) Predicted activation z-scores were combined with actual fold-change (log2) of top differentially expressed genes across the stage-specific datasets including: (C) initiation in non-diabetic lesions −/−, (D) progression in non-diabetic lesions −/−, (E) all non-diabetic lesions, (F) moderate lesions in diabetics and non-diabetics (+/−), (G) moderate lesions in diabetics and non-diabetics (+/−), and (H) all diabetic lesions. Circos plots for transcription factors that are dynamically regulated only in (I) non-diabetic atherosclerotic lesions or (J) diabetic lesions.

Because IPA upstream regulators may not directly be involved in transcriptional regulation, we next focused on identifying differentially expressed transcription factors (Reads>100, log2FC(0.5), FDR<0.1) using TFcheckpoint and Panther Gene List Analysis.^27,28 Stage-specific transcription factors for non-diabetic and diabetic progression of disease were summarized in circus plots, displaying their expression in relation to the stage that they are most commonly associated (Suppl. Figure VA). Identified transcription factors in the non-diabetic comparisons such as FHL2, ID3, and FOXP2 are significantly increased with progression of atherosclerosis, but do not further increase or decrease in diabetic lesions (Suppl. Figure VB–E). Commonly regulated transcription factors could serve as important checkpoint genes for triggering the onset of the next stage. For example the transcription factor lipopolysaccharide-induced TNF-a factor (LITAF), a gene that regulates CCL2 and STAT6B ²⁹ but has not been described in the context of atherosclerosis, is increased with progression of atherosclerosis and even more with diabetes (Figure 5I). Compared to non-diabetic lesion progression, the two highest expressed transcription factors in the onset of diabetic atherosclerosis were HIF1A and HEYL, with HIF1A being increased with diabetes and HEYL decreased with diabetes compared to non-diabetic groups (Figure 5J). The identified transcriptional upstream regulators in this section may explain observed downstream GO processes as a result of a signaling cascade. Therapeutically targeting these upstream nodes, such as CXCR4, may affect commonly activated PI3K processes.

Validation study of selected genes by NanoString profiling.

The discovery genome-wide RNA-Seq studies lead to the identification of a range of differentially expressed genes, signaling pathways, hubs, upstream regulators, transcription factors, and biological processes. To investigate the accuracy of this discovery set, we performed a validation study. We examined the power of 55 selected genes from each of the aforementioned discovery categories that were specific to either a stage of disease progression or commonly regulated in diabetes or non-diabetes conditions to create a customized NanoString chip. Unsupervised hierarchical heatmap cluster analysis of 55 selected genes is shown in Figure 6A. NanoString data were normalized to negative control genes (n=5) that were not differentially expressed in the initial RNA-Seq. Pearson correlation was carried out for all comparisons. For example, the identified genes derived from the initiation of atherosclerosis group in non-diabetic lesions (i.e. initiation −/−) correlated with the fold-change retrieved from NanoString analysis (R=0.96, p<0.0001) (Figure 6B). In summary, the 55 identified genes from RNA-Seq studies derived from all 23 RNA lesion-based samples highly correlated in validation studies obtained from NanoString profiling (Figure 6C, Suppl. Figure VI). Unsupervised heatmap cluster analysis shows the accuracy of the initial RNA-Seq studies as well as how distinct the non-diabetic (orange) expression profile is compared to diabetic-derived (blue) lesional RNA (Figure 6D). In addition, PCA of only 55 genes is able to discriminate non-diabetic and diabetic samples (Figure 6E). In summary, NanoString profiling confirmed the accuracy of fold-changes of these genes from the discovery set identified by RNA-Seq. Moreover, PCA of only 55 selected genes is able to cluster the samples in a similar manner from the RNA-Seq analyses shown in Figure 2B, where all genes were included.

Figure 6. Validation study of RNA-Seq samples by NanoString profiling.

(A) Heatmap cluster analysis of 55 selected genes using RNA-Seq coverage values. (B) Pearson correlation of 55 genes comparing RNA-Seq versus NanoString profiling from the group initiation in non-diabetic lesions −/−. (C) Correlation studies and (D) heatmap cluster analysis across all group comparisons from RNA-Seq to NanoString profiling. (E) Principal component analysis (PCA) of NanoString for 55 genes using all 23 samples of the initial RNA-Seq study.

Translational value of identified genes on human specimen samples.

To assess the translational value of these 55 genes in human specimens, we applied the same NanoString chip to a cohort of human atherosclerotic or control carotid artery specimens with or without diabetes. We analyzed the same 55 genes on NanoString as described above from RNA derived from carotid artery cross sections of healthy/control (n=4) or patients with clinical manifestation of carotid atherosclerotic disease (n=16) of which 31% had type 2 diabetes mellitus (T2DM), 13% diet-controlled T2DM, and 50% without diabetes (Figure 7A, Suppl. Table III). 76% of probes were conserved from pig to human. Data was normalized using edgeR and non-differentially expressed genes were removed for unsupervised and supervised cluster analysis (Figure 7A). As a first step, unsupervised cluster analysis showed clear separation of control non-atherosclerotic subjects to atherosclerotic patients. The gene profile of lesions from diet-controlled T2DM patients (n=2) resembled the non-diabetic profile, and separated from the T2DM patients. Indeed, we observed two distinct gene profiles for these T2DM samples (Figure 7B,C). Correlation with clinical available data showed that hemoglobin A1c, insulin, platelets (Plt), low density lipoprotein (LDL), and triglyceride (TG) correlated with severity (i.e. normal = 1, non-diabetic = 2, T2DM = 3), but not with other factors such as age, body mass index (BMI), cholesterol. (Suppl. Figure VIIA,B). With addition of clinical data (i.e. supervised), PCA separation improved for non-diabetic and T2DM atherosclerotic samples (Figure 7D, Suppl. Figure VIIC,D).

Figure 7. Validation of NanoString profiling on human atherosclerotic and control specimens with or without type 2 diabetes mellitus (T2DM).

(A) Workflow for human specimen cohort. (B,C) Unsupervised cluster analysis for n=20 samples based on their expression of 26 genes. (D) Supervised cluster analysis including patient history data in combination with expression of 26 genes.

Identification of a three-mRNA signature for different states of hypercholesteremia with and without hyperglycemia.

Finally, machine learning cluster analysis was performed on the human NanoString data set. This led to the identification of F8, MAPKAPK3, and ITGB1 which separated healthy control from atherosclerotic specimens based on their absolute expression (Figure 8A). Remarkably, these three genes also clustered groups into T2DM, diet-controlled T2DM, and non-diabetes compared to healthy control subjects as shown in three-dimensional (3D) cluster plots (Figure 8B). This segregation across different disease stages was even more pronounced when removing the control group (Figure 8C). These three genes (i.e. F8, MAPKAPK3, and ITGB1) were further validated on the initial Yorkshire RNA-Seq data set. Indeed, the three genes had sufficient power to cluster the two diabetic groups (moderate and severe) from the three non-diabetic groups (mild, moderate, severe) (Figure 8D). Moreover, the three genes could also distinguish the three non-diabetic groups for their degree of severity (Figure 8E). Finally, the expression of F8, MAPKAPK3, and ITGB1 clustered the RNA-Seq data in all its individual groups, where the two diabetic groups are closer compared to the three non-diabetic groups (Figure 8F). Protein expression was examined on swine tissue slides for F8, MAPKAPK3, and ITGB1. F8 protein expression correlated with transcriptomics: the expression increased in non-diabetic coronary arteries and decreased in diabetic coronary arteries (Suppl. Figure VIIIA). MAPKAPK3 on the other hand increased significantly on protein expression in non-diabetic severe lesions, but not in diabetic lesion, which agrees with the findings from the transcriptomics data (Suppl. Figure VIIIB). For ITGB1, we observed the expected reduction in expression in non-diabetic moderate pigs; however, there was no change in ITGB1 expression in the diabetic lesions (Suppl. Figure VIIIC).

Figure 8. Three-Dimensional representation of clusters in human and swine lesion cohorts based on expression of ITGB1, F8, and MAPKAPK3 identified by machine learning.

The 3D plots are shown representing the separation of: (A) human control groups (i.e. no atherosclerotic lesions) compared to those patients with atherosclerotic lesions based on expression levels of the indicated genes using NanoString; (B) Human patients with atherosclerotic lesions with T2DM, diet-controlled T2DM, or no-T2DM compared to controls without atherosclerotic disease using NanoString. (C) Atherosclerotic lesions with T2DM, diet-controlled T2DM, or without T2DM. (D) Atherosclerotic lesions from swine with and without diabetes based of the expression levels of the indicated genes using RNA sequencing platform. (E) Separation of different clusters across increasing severity of lesions in swine in non-diabetics based on the expression levels of the indicated genes using the RNA sequencing platform. (F) Separation of different clusters across increasing severity of lesions with or without diabetes based of the expression levels of the indicated genes using the RNA sequencing platform.

Collectively, the original discovery set from genome-wide RNA-Seq of pig lesions was successfully validated in human atherosclerotic and control subjects with and without diabetes. Although the human cohort was limited, distinct groups for the control, non-T2DM and T2DM occurred. The power of machine learning facilitated narrowing down this set to only three genes, which not only bifurcated non-diabetic vs. diabetic lesions, but also separated increasing severity of atherosclerotic lesions.

DISCUSSION

Atherothrombotic cardiovascular disease leads as a cause of death worldwide and is exacerbated by the growing epidemic of obesity-induced insulin resistance and diabetes. Despite lifestyle preventive approaches and optimal medical therapies, residual cardiovascular risk remains. A better understanding of how diabetes promotes atherosclerosis from initiation to progression may facilitate new targets for this devastating disease. The use of aortic atherosclerotic preparations from rodents has several limitations including differences in plaque composition compared to coronary arteries of larger animals or humans. Identification of lesion-based transcriptomic profiles from coronary arteries of pigs may provide important insights into the development of both diabetic and non-diabetic coronary atherosclerosis.

We tested the hypothesis that transcriptional diversity is more pronounced than histopathological markers to account for lesional differences with the imposition of diabetes in laboratory swine. Therefore, this study was designed to perform an RNA-Seq discovery study of lesion-derived RNAs, which were characterized initially for their phenotypic histopathological manifestations (i.e. I/M-ratio, ORO staining, CD45, elastin) in non-diabetic and diabetic coronary arteries (Figure 1). Multiple lesions were characterized in each animal, hence this study was designed lesion-centric and not from independent animals. As do humans, coronary artery atheroma in pigs display a range of lesion types. The histopathological characterization for severity of atherosclerosis into mild, moderate, and severe progression of disease was assessed after 60 weeks on HCD. In support for accelerated atherosclerosis with diabetes, diabetic pigs contained ~2.5-fold greater proportion of plaques compared to the non-diabetic group. Furthermore, while lesions were detected as mild, moderate, and severe in the non-diabetic group, only moderate to severe lesions were found in the diabetes group consistent with accelerated atherosclerosis. Among the non-diabetic groups, moderate and severe RNA-Seq groups were compared each to mild, which reflects the progression of atherosclerosis (Figure 2). Not surprisingly, inflammatory-associated pathways were enriched at early stages (moderate vs. mild), whereas with progression oxidative stress, mitochondrial dysfunction, and LXR/RXR pathways became dominant (Figure 3). Specific signaling pathways emerged for glycoprotein VI platelet (GP6), sirtuin, RhoA, and protein kinase A as activated nodes only under non-diabetic conditions (Figure 4A). The transmembrane domain of GP6 is constitutively associated with the Fc receptor γ (FcRγ), which is essential for its function as a downstream immunoreceptor tyrosine-based activation motif (ITAM).³⁰ GP6 is a promising pharmacological target for the treatment of thrombosis as blocking the interaction to its ligand collagen by a soluble dimeric GP6-Fc fusion protein reduced thrombosis in mouse models and is currently in clinical trial (clinicaltrials.gov: NCT01645306).³¹ Other hubs revealed new targets involved in tyrosine kinase signaling or adaptor proteins. For example, the transmembrane signaling enzyme PLCG2 emerged as one of the top ten most frequently involved hubs only in the non-diabetic groups (Figure 3F). Among the top identified upstream regulators in the non-diabetic groups was SPP1, also known as osteopontin or leukocyte activation 1, which is induced in the non-diabetic groups from initiation to progression of lesions, but not in diabetic lesions (Figure 5C). Osteopontin deficiency attenuated atherosclerosis in ApoE^−/− mice,³² suggesting its relevance in regulating early inflammatory responses. Another top gene OLR1, also known as LOX-1, is expressed with progression of non-diabetic lesions, and a polymorphism in its locus is associated with increased risk for atherosclerosis-related diseases.³³ Collectively, the transcriptomic analysis revealed stage-specific processes with inflammatory signaling dominating the initiation stage, whereas oxidative stress and mitochondrial dysfunction are more prominent with lesion progression in non-diabetics.

To identify gene signatures specific for diabetic atherosclerosis progression, similar stage lesions from diabetic-rendered pigs were chosen for non-diabetic lesions (Figure 1). Although histopathological markers such as ORO, CD45, or I/M ratio differed little if at all in the selected diabetic lesions compared to non-diabetic moderate and severe lesions, we observed substantial transcriptomic changes (Figure 2A). Inflammatory Th17 and IL-17F signaling pathways were enriched specifically in diabetic lesions. While IL-17 signaling has been reported to have both pro- and anti-atherogenic roles in non-diabetic mice (30, 31), IL-17 family members directly compromise β-cell function and hence promote the development of diabetes.³⁴ Moreover, inhibition of Th17 cells in mice due to IL-17 deficiency mice led to suppression of diabetes.³⁵ These reports suggest that IL-17 signaling may accelerate atherosclerosis by exacerbating diabetic conditions. We also found that the two highest differentially expressed transcription factors in the onset of diabetic atherosclerosis were HIF1A and HEYL (Figure 5J), with HIF1A being increased with diabetes and HEYL decreased with diabetes compared to non-diabetic groups. Interestingly, diabetic tissues have been described to express higher levels of HIF1A as hyperglycemia often associates with a phenomenon called pseudohypoxia, where the increased glucose flux disturbs the NADH/NAD⁺ ratio leading to metabolic stress.^36,37 Reduced sirtuin signaling in diabetic lesions is also consistent with predisposition to DNA damage, vascular senescence, impaired NADH/NAD⁺ ratio, and mitochondrial dysfunction.³⁸ HEYL expression correlates with endothelial cell aging and is involved in regulation of the Notch pathway, suggesting the importance of aging-related endothelial cell senescence for the progression of diabetes.³⁹ These diabetic gene signatures demonstrate the value of unbiased genome-wide transcriptomic analysis to identify canonical pathways and unique gene signatures and provide a strong foundation to dissect on the molecular level distinguishing features in non-diabetic and diabetic-associated atherosclerosis.

The accuracy of the initial RNA-Seq was validated using a customized NanoString chip with 55 selected genes including differentially expressed genes, signaling pathways, hubs, upstream regulators, and transcription factors. We not only observed the same fold changes by RNA-Seq, but also were able to use only the expression profile of those 55 genes to obtain the same clustering for all the five groups. To test the translational value of our study to human lesions, the expression of the selected genes for NanoString was evaluated on a small, but well-defined set of carotid artery atherosclerotic specimens with and without diabetes. Out of the 55 genes on the NanoString, 26 were differentially expressed and both unsupervised and supervised cluster analysis showed a distinct pattern for control and atherosclerotic lesions from non-diabetic subjects (Figure 7B). Moreover, PCA clearly separated non-diabetic and T2DM atherosclerotic samples (Figure 7D, Suppl. Figure VI C,D). Interestingly, the diet-controlled diabetic samples clustered with the non-diabetic group, which might be expected as those patients had well-controlled diabetes based on their HbA1c levels. Thus, the separation of clusters using this small set of genes supports the premise of a conserved gene signature of diabetic and non-diabetic lesions from pigs to humans.

Surprisingly, machine learning identified that among the differentially expressed genes (n=26) in the human cohort a three-mRNA signature was able to discriminate the different disease states. Based on the expression of ITGB1, MAPKAPK3, and F8, all groups in the human cohort could be separated depending on the presence of atherosclerosis and whether they were T2DM or not (Figure 8A–C). Although this three-mRNA signature was identified based on human expression data, it could also cluster all the RNA-Seq groups respectively (Figure 8D–F). ITGB1, a membrane receptor protein involved in cell adhesion, has been linked to diabetes as mice lacking ITGB1 show reduced expansion of pancreatic β-cells.⁴⁰ F8, the coagulation factor VIII, has been linked to atherothrombosis. For example, ApoE^−/− mice deficient for F8 were atheroprotective, as reduced hypercoagulability provides protection against acute thrombus formation.⁴¹ Indeed, patients with diabetes exhibit significantly elevated plasma F8 levels and increased fibrinogen concentration compared with healthy subjects.⁴² Lastly, MAPKAPK3, also known as MK-3, is a member of the Ser/Thr protein kinase family and shows 75% sequence homology to its paralog MK-2.⁴³ Systemic knockout of MK-2 in Ldlr^−/− mice developed reduced atherosclerosis compared to control.⁴⁴ Recent work has demonstrated that an allosteric MK-2/MK-3 inhibitor in combination with the drug metformin cooperatively improved glucose levels through suppressing hepatic glucose production.⁴⁵

Future studies will be of interest to delineate the functional roles for each of these top regulators in diabetes-associated atherosclerosis. Moreover, these gene signatures have particular relevance for the multistep process involved in atherosclerosis that evolves over time from a fatty streak to advanced lesions prone to plaque rupture, erosion, or quiescence. To accommodate changes in the molecular composition and interactions, identification of these complex spatial and temporal transcriptomic relationships involves a trade-off between ease of expression profiling and physiological relevance.⁴⁶ Hence, this study overcomes those challenges often found using aortic preparations from athero-prone mice ^14,47 by using coronary arteries from pigs as a more pathophysiologically relevant model of atherosclerotic disease. As diabetes exacerbates atherosclerosis and the number of human subjects with metabolic syndrome, pre-diabetes, and diabetes continues to grow,⁴ insights into the gene regulators and upstream drivers of diabetes-associated atherosclerosis may provide opportunities for new targets for therapy or diagnostics of this accelerated disease. This point is underscored by the higher rates of major adverse cardiac events that remain in diabetic subjects despite optimal therapeutic strategies that focus on anti-lipid, anti-hypertensive, and glucose lowering treatments.⁴⁸

Several limitations of this study merit consideration. First, the time point used to harvest the pig coronary artery lesions was after 60 weeks of hypercholesterolemia or hypercholesterolemia with diabetes. Therefore, we cannot exclude the possibility that other gene drivers may be more relevant if harvested at earlier time points – though at these early times lesion formation would be challenging to discern. Our focus on analysis of lesions of varied severity across animals for mild, moderate, and severe manifestations of disease should yield stage-specific results independent of time points harvested. Moreover, we recognize that RNA was extracted from whole cross sections, including the tunica adventitia, which may contribute to the observed inflammatory signaling footprint. Another consideration is potential differences that may exist related to the induction of diabetes using STZ in the swine compared to the diabetes-associated atherosclerosis in the analyzed cohort of human subjects. A single injection of STZ offers the advantage of producing and maintaining diabetes for prolonged periods in pigs. The diabetic, hyperlipidemic swine used herein, initially described by Gerrity et al.,¹² exhibit features of both type 1 and type 2 diabetes due to the lack of insulin dependence, hypertriglyceridemia, and glucose intolerance and has served extensively for over two decades to reveal new histopathological, hemodynamic, and molecular insights of accelerated coronary artery lesions compared to non-diabetes, hyperlipidemic animals. Although some mechanisms differ in driving atherosclerotic lesions in type 1 and type 2 diabetes, both associate with accelerated atherosclerosis.⁴⁹ Diabetes is an independent risk factor for atherosclerosis regardless of the type, duration, or severity.⁵⁰ Nonetheless, future studies should explore the markers identified here in other human cohorts with different types of diabetes and atherosclerotic disease. Male swine were used in this study to recapitulate the hypercholesterolemic and diabetic conditions achieved as previously described¹² and though we recently reported⁵¹ no sex differences in vascular lesion formation in swine, our current study may not reflect similar phenotypic and transcriptomic patterns in female swine of similar age. Future studies will be of interest to assess for sex-specific differences. Single cell RNA sequencing could provide more information regarding gene expression in particular cell populations. Yet, in the complex and extracellular matrix-rich lesions in pigs and humans, enzymatic and mechanical dissociation protocols appropriate for thin mouse arteries can introduce considerable bias in single cell yields. Too mild a dissociation protocol leaves many cell clumps, while a more aggressive protocol can lyse cells precluding their analysis. We recognize that other external factors may influence total atherosclerotic burden that could in theory impact lesion composition, including the presence of diabetes. Nevertheless, we did not observe increased frequency of lesions in diabetic coronary arteries, thereby minimizing this as a confounder. Moreover, pig coronary lesions and human carotid lesions may harbor distinct transcriptomic signatures; future studies will be of interest to compare broad transcriptomics profiling of different human and pig vascular beds. Finally, future studies will be of interest to compare single cell RNA-seq with bulk RNA-seq to provide additional insights for cell-specific transcriptomic profiles.

In summary, using a lesion-based transcriptomic profiling approach, this study identified specific and common gene signatures in coronary arteries of hypercholesterolemic pigs with or without diabetes. Despite similar histopathological features between non-diabetic and diabetic lesions, we found distinct gene signatures, upstream regulators, canonical signaling pathways, and cellular processes that establish a new foundation for future translational investigation. We further identified three genes (i.e. F8, MAPKAPK3, and ITGB1) that are able to distinguish diabetes and non-diabetes-associated atherosclerosis in both pigs and humans. These findings provide new insights for gene pathways and therapeutic opportunities to potentially impact the accelerated atherosclerotic lesion formation in diabetes.

HIGHLIGHTS.

• This lesion-centric study profiles non-diabetic versus diabetic gene signatures with increasing severity of lesions in the coronary arteries of swine.

• Despite similar histopathological characteristics of lesions from animals with or without diabetes, the transcriptomic profiles were profoundly different.

• We identify and validate network hubs that drive key signaling pathways, transcription factors, and canonical pathways in a stage-specific manner in diabetic and non-diabetic lesions.

• Machine learning algorithms identified three genes (i.e. F8, MAPKAPK3, and ITGB1) that distinguished diabetes and non-diabetes-associated atherosclerosis across pigs to human subjects.

Abbreviations and Acronyms

ApoE

Apolipoprotein E

DM

Diabetes Mellitus

PCA

Principal component analysis

CVD

Cardiovascular Disease

HCD

High Cholesterol Diet

IPA

Ingenuity Pathway Analysis

I/M-Ratio

Intima/Media-Ratio

Ldlr

Low Density Lipoprotein Receptor

ORO

Oil Red O Staining

SMC

Smooth muscle cells

STZ

Streptozotocin

T2DM

Type 2 diabetes melilites