Virtual Transcriptomics: Non-invasive Phenotyping of Atherosclerosis by Decoding Plaque Biology from Computed Tomography Angiography Imaging

Abstract

Objective:

Therapeutic advancements in atherosclerotic cardiovascular disease have improved prevention of ischemic stroke and myocardial infarction, but diagnostic methods for atherosclerotic plaque phenotyping to aid individualized therapy are lacking. In this feasibility study, we aimed to elucidate plaque biology by decoding the molecular phenotype of plaques through analysis of computed-tomography angiography (CTA) images, making a predictive model for plaque biology referred to as virtual transcriptomics.

Approach and Results:

We employed machine intelligence using paired CTA and transcriptomics from carotid endarterectomies of 40 patients undergoing stroke-preventive surgery for carotid stenosis. CTAs were analysed with novel software for accurate characterization of plaque morphology and plaque transcriptomes obtained from microarrays, followed by mathematical modelling for prediction of molecular signatures. 414 coding and non-coding RNAs were robustly predicted using supervised models to estimate gene expression based on plaque morphology. Examples of predicted transcripts included ion transporters, cytokine receptors, and a number of microRNAs whereas pathway analyses demonstrated enrichment of several biological processes relevant for the pathophysiology of atherosclerosis and plaque instability. Finally, the ability of the models to predict plaque gene expression was demonstrated using CTAs from four sequestered patients and comparisons with transcriptomes of corresponding lesions.

Conclusions:

The results of this pilot study show that atherosclerotic plaque phenotyping by image analysis of conventional CTA can elucidate the molecular signature of atherosclerotic lesions in a multi-scale setting. The study holds promise for optimized personalized therapy in the prevention of myocardial infarction and ischemic stroke, which warrants further investigations in larger cohorts.

Graphical Abstract

Introduction

Cardiovascular disease (CVD) is the most common cause of death and disability in the world, mainly by myocardial infarction and ischemic stroke from unstable atherosclerosis,¹ which exerts an exorbitantly high financial burden to society.² Risk management of patients is largely dependent on population-based scoring methods such as the Framingham Risk Score or secondary prevention in patients with established disease^{3, 4} and development of diagnostics for more precise patient categorization is warranted. Despite discoveries of new predictive plasma biomarkers⁵ and improved plaque imaging,⁶ routine diagnostic methods for identification of individuals and lesions at high risk for atherothrombosis in coronary or extracranial arteries are still lacking.⁷ In addition, strategies to implement tailored, personalized pharmacotherapy remain limited without practical non-invasive assessment of biological and molecular disease features.⁸

Diagnostic accuracy would improve if the morphological and biological features of atherosclerotic plaques could be determined by non-invasive imaging. In order to do this, a quantitative linkage must be established between scales. Mathematical formalisms for such multi-scale models have been explored,^9-16 such as intracellular cell cycle networks (e.g., by ordinary differential equations; ODEs), regulation of the collective behaviour of cellular structures (e.g., by cellular Potts models; CPM), spatially differentiated behavior (by partial differential equations; PDEs), which in turn control intracellular behavior (returning to ODEs). Proposed model creation and validation workflows^17-25 rely on high level hypothesis generation and then collection of data. Whereas data has been documented at individual scales, i.e., in vitro data at cellular and molecular scale, microscopy data at histopathology scale, and radiological data at macroscopic scale, there is a dearth of linkages across these scales.

Development of quantitative imaging biomarkers (QIBs) for guiding cancer therapy based on non-invasive imaging using molecular signatures from tissue biopsies as a truth basis, has been met with enthusiasm.^{26, 27} A similar approach is more challenging for coronary artery disease (CAD) as acquisition of plaque tissue biopsies from living patients is not generally possible. However, combined analyses of preoperative imaging and lesions can be performed in patients with carotid atherosclerosis undergoing stroke-preventive surgery (carotid endarterectomy; CEA).²⁸ Carotid atherosclerosis shares morphological and biological features with CAD and may thus serve as an attractive disease model in development of QIBs for unstable atherosclerosis in other vascular beds.²⁹ Investigations in animal models and histological analyses of human lesions have characterized distinct but common structural and biological features such as enhanced inflammation, accumulation of a large lipid-rich and necrotic central core (LRNC), intra-plaque haemorrhage (IPH), a thin and rupture-prone fibrous cap from extracellular matrix (ECM) degradation and apoptosis of smooth muscle cells (SMCs), and generally less calcified (CALC) than more stable, asymptomatic, counterparts.³⁰ More recently, the morphological and biological features of atherosclerotic plaques in humans have also been corroborated by molecular pathway analyses of the human plaque transcriptome.^{31, 32}

In this pilot study, we aimed to decode atherosclerotic plaque molecular phenotype non-invasively, making a predictive model for gene expression that we refer to as virtual transcriptomics. Our approach was focused on training machine intelligence models to interpret conventional CTAs with paired global microarray-based transcriptomic analyses of CEAs utilizing an established human biobank (BiKE; Biobank of Karolinska Endarterectomies).³³ The study demonstrates the feasibility of using non-invasive, commonly available imaging protocols combined with advanced morphological and molecular characterization of atherosclerotic plaques and machine intelligence methods to determine per-patient molecular level signatures, with potential for optimizing personalized therapy in the prevention of myocardial infarction and ischemic stroke.

Methods

Transcriptomics data is available from Gene Expression Omnibus (GSE125771) and other data in compliance with ethical restrictions are available from the corresponding author upon reasonable request.

Study Workflow

The study used a novel image analysis technique for atherosclerotic lesions in CTAs together with gene expression by microarrays as an objective truth basis of plaque biology to create computational models for predicting molecular plaque signatures, determine plaque phenotype and aid clinical decision making in patients without analysis of tissue specimens (Fig. 1). Two levels of ground truth were used, one to support characterization of plaque tissue by CTA based on histology from an independent tissue bank³⁴ to create “virtual tissue models” (Fig. 2 A-E; Supplement Fig. VIII), and a second one to quantify molecular mechanisms based on transcriptomics.³⁵ All data has been corrected for age and sex. Resulting models were then deployed on previously unseen, non-invasive imaging data (hold-out patients) for which actual transcriptomics was available for validation. This concept provided a hypothesis of predictive “virtual transcriptomics” capability on a per-patient basis, tested in this study.

Figure 1:

Study work flow showing plaque characterization using two levels of machine intelligence models to provide both quantitative plaque morphology as well as estimated tissue gene expression on a per-patient basis. Measurements made at the first level of processing included structural anatomy and tissue characterization, using the vascuCAP software’s Virtual Tissue Models, which were trained using pathologist annotated specimens. These were then fed forward as inputs to the models to elucidate molecular profiles determining plaque phenotype. The experimental workflow utilized a set of cases with paired transcriptomic data from microarrays in a development cohort and subsequently in a cohort of sequestered test patients. This truth data was used to build the virtual expression models in the development cohort, then locked down for application to the held-out test patients as a validation of model capability. Purple shades annotate quantitative plaque morphology where dark signifies a high value; red/green scale is used for expression, where green signifies high expression, red low, and black signifies intermediate. Unless otherwise noted in figure legends or text, the scales are log transformed but graph axes are denoted in original scale.

Figure 2:

Multiple objectively validated measurements and characterizations were made to characterize plaque morphology by CTA analysis software. These assessments included structural anatomy (“structure”) and tissue characterization (“composition”). (LRNC=lipid-rich necrotic core, IPH=intra-plaque haemorrhage, PVAT=perivascular adipose tissue). Analysis of carotid plaque tissue composition in the bifurcation of the left carotid artery in an 86 y.o. white man with a transient ischemic attack: Processed 3-D image of the artery and lesion from the CTA (A; common carotid artery partition – red; internal carotid - chartreuse; external carotid artery - purple); Histological section of the CEA specimen just distal of the bifurcation (B); section annotated by a pathologist indicating presentation of CALC (green) and LRNC (yellow). Cross sectional CTA image positioned near the histological section with outer wall and lumen segmented by the software with tissue characterization suppressed to show raw imagery (C) (here, MATX is shown with two colours to reflect that the pathologist marked dense fibrosis in dark blue and remaining MATX elements are indicated as a dark grey); the wall outline on CTA (D); the software’s characterization of tissue composition (E). Note that only those tissue types present in a given sample appear; in this example, IPH and PVAT are not present. Colours in panels A-E: yellow - LRNC; aquamarine – CALC; light green - outer vessel wall boundary; orange - lumen boundary; dark blue - fibrotic tissue; blue - MATX.

Human Samples and Plaque Tissue Transcriptomics

A total of 44 patients (40 development, 4 sequestered test) undergoing stroke-preventive CEA for high-grade (> 50% NASCET³⁶) carotid stenosis were used in this study. Patients with high vs. low calcified carotid lesions on CTA were selected as previously described³⁷ and the study cohort demographics summarized in Supplement (Table I). Briefly, CEAs were collected at surgery and retained within the biobank (BiKE), details of sample collection and processing, and transcriptomic analyses by Affymetrix microarrays were as previously described.^{38, 39} Briefly, plaques were divided transversally at the most stenotic part; the proximal half of the lesion was used for RNA preparation while the distal half was fixed in 4% formaldehyde and prepared for histology. The microarray dataset is available from Gene Expression Omnibus (GSE125771). All samples were collected with informed consent from patients and the study was approved by the Ethical Review Board.

CTA Image Analysis

Carotid CTA exams from the aortic arch to the vertex were performed with 100 or 120 kVp, variation of CTDIvol16cm between 13.9-36.9 mGy or CTDIvol32cm 7.9-28.3 mGy with intravenous contrast administered as previously described.³⁷ Axial image reconstruction of 0.625 mm were obtained and transferred for image analysis performed by E.K, blinded to histological and biochemical analysis. The vascuCAP^® (Elucid Bioimaging Inc., Boston, MA) software^{32, 34, 40-43} was used to provide characterization of plaque morphology (Fig. 2; Supplement Fig. VIII). The software creates fully 3D segmentations of lumen, wall, and each tissue type at an effective resolution approximately 3x higher than the reconstructed voxel size with improved soft tissue plaque component differentiation relative to manual inspection. The common and internal carotid artery were defined as a target with lumen and wall evaluated automatically and, when needed, edited manually. The external carotid artery was excluded and image analysis limited to the proximal half of the lesion, corresponding to the tissue used for RNA isolation and microarray analysis.

The vessel wall was analyzed defining the plaque into different components: LRNC, CALC, IPH, matrix (MATX; representing plaque tissue not belonging to the other types), perivascular adipose tissue (PVAT), cap thickness (the smallest distance from LRNC to the lumen), and degree of stenosis. The software includes algorithms to decrease blur caused by image formation in the scanner. A patient-specific 3D point spread function is adaptively determined so that image intensities are restored to more closely represent the original materials imaged, which mitigates artefacts such as calcium blooming, and enables discrimination of less prominent tissue types.

The image restoration is undertaken in concert with a novel method for tissue characterization based on expert-annotated histology (Fig. 2; Supplement Fig. VIII). The overlapping densities of tissues such as LRNC and IPH necessitated a method for accurate classification. To avoid limitations of conventional analysis of CTA utilizing fixed thresholds, the accuracy required for elucidating molecular pathways was achieved by algorithms that account for distributions of tissue constituents rather than assuming constant material density ranges. In this way, the software makes mathematical judgements to interpret the Hounsfield units (HU) of adjacent voxels by maximizing criteria that mimic expert annotation at microscopy, simultaneously mitigating variation between scanners, reconstruction kernels, and contrast levels. In this way, the software fundamentally addresses subjectivity intrinsic to other analysis methods.

Analytic performance of the software has been undertaken both for tissue composition accuracy relative to histopathology³⁴ and reader repeatability and reproducibility.⁴⁴ Performance characteristics are included in the commercially available clinical edition software regulated as a medical device.⁴⁵

Predictive Modelling

Of the 54,676 probes for coding and non-coding RNAs represented in the microarray, 3478 probes were selected as most relevant to atherosclerosis based on the following criteria. First, we selected genes that were found to be highly dysregulated in comparisons of lesions with differing levels of calcification.³⁷ Briefly, global gene expression analysis comparing high vs. low calcified plaques (30-65% of plaque area vs. 0-2%) resulted in 3387 significantly differentially expressed probe-sets, of which 1783 were upregulated and 1604 downregulated (of total 70526 microarray probe sets, Bonferroni adjusted p < 0.05). We then selected transcripts previously documented as being dysregulated in plaque instability³³ as well as those identified in a systems biology survey of atherosclerotic mechanisms,⁴⁶ adding a net of 91 additional transcripts identified in the cited works by symbol that were not already contained in the experimentally determined 3387.

Single as well as multiple variable regression models covering linear and non-linear modelling techniques were performed on predictor sets constructed from the development cohort (n=40) including plaque morphology, demographics, clinical (laboratory) values, and stenosis. In part to recognize that clinical factors could have affected the expression data or models, and to inspect what is the added value of morphology over clinical and demographic data, we modelled these effects using different predictor sets, some only using plaque morphology, but others also using lab values, demographic, and other values in composite models. Each model result was output to tabulate the highest-achieved performance on a transcript-by-transcript basis. Predictive performance was determined based on the accuracy of the prediction relative to the true value in each of the 3478 transcripts.

All models were built with three levels of variation: (i) differing sets of morphological measurements according to hypothesized physiological rationale (on all 3478 transcripts); (ii) automated optimization using 10-fold cross validation while simultaneously varying tuning parameter values (on all 3478 transcripts); and (iii) data was partitioned such that a training set on which the cross-validation was performed was strictly separated from a sequestered validation data set to test performance using locked-down models. Use of histologically validated plaque features produces interpretable models,⁴⁷ coupled with cross-validation, mitigated overfitting.

Statistical Analysis

Supervised and unsupervised methods were applied to assess the ability of CTA morphological measurements to identify molecular mechanisms obtained from transcriptomics of paired CEA specimens. Unsupervised clustering was used to provide a rough sense for relationships between plaque morphology and expression levels. The hierarchical clustering is represented as a dendrogram split at points with Pearson correlation less than 0.8 using a Euclidean distance function according to the complete linkage method on both plaque morphology measurement features and on expression levels, plotted as a heatmap.

Supervised model quality (MQ) was determined as the product of two measures for each model type. MQ for continuous estimation models was computed as the product of concordance correlation coefficient (CCC) and regression slope of predicted vs. observed for continuous value estimation (the former to measure the tightness of fit, but augmented by the latter to ensure proportional prediction relative to observed). MQ for dichotomized categoric prediction models was computed as the product of area under the receiver characteristic curve (AUC) times Kappa for dichotomized prediction (the former to measure the net classification performance, but augmented by the latter to ensure performance in both high and low expression classes). Transcripts were classified as “robustly predicted” if dichotomized MQ exceeded .15 (e.g. as met by AUC of .75 and Kappa .2), and were included in unsupervised clustering analyses. Those with MQ exceeding .4 (e.g. as met by AUC of .8 and Kappa .5) were classified as “particularly robustly predicted” and further analysed by gene-set enrichment analysis (GSEA) to elucidate biological processes and molecular pathways at the cohort level, as well as being included in test patient validation. GSEA was conducted using EnrichR (https://amp.pharm.mssm.edu/Enrichr/), further passing results from Gene Ontology Biological process 2018 with p- <0.05 values (adjusted for multi-hypothesis testing) to Revigo (http://revigo.irb.hr/) to determine non-duplicative processes, and finally merged with Reactome 2016 pathways that fell in the same range of significance.

Models were then fixed (“locked down”) and applied to a sequestered set of patients (n=4) selected at random for which ground truth was known, to validate the performance of the model on patients not included in development of the model (“unseen patients”) to test generalizability.⁴⁸ For each test patient, we used the models for transcripts that were particularly robustly predicted and determined the significance of the predictions by applying a bootstrap method to permute plaque morphology inputs to each transcript’s model, providing a measure of model stability used to adjust the outputs for each test patient. Model predictions were sorted by individualized confidence and the top 20 most significantly dysregulated transcripts plotted (ranked by combining the degree of dysregulation and the statistical significance in its estimation), for each patient, which was finally compared with the true expression of corresponding transcripts. We then proceeded to pathway analysis by GSEA using the particularly robustly predicted transcripts for each patient to provide a patient-specific unbiased determination of dominant mechanisms. The patient-specific GSEAs were determined from transcript ranking (see above), and p-values for the process were adjusted for multiple hypothesis testing.

Results

Unsupervised Relationships between Plaque Morphology and Gene Expression

Plaque morphology from image analysis of CTAs in 40 patients tested against gene expression levels for 3478 selected transcripts generated 414 transcripts meeting the MQ criteria for robustly predicted by plaque morphology, and subsequently subjected to unsupervised clustering (Fig. 3 and Supplement Table V). High CALC was associated with high expression levels of Proteoglycan 4 (PRG4) and low levels of Speedy/RINGO Cell Cycle Regulator Family Member E1 (SPDYE1), as examples (Supplement Table VI). High LRNC, was for example coupled to high expression of Matrix Metalloproteinase 12 (MMP12) and low levels of Rap Guanine Nucleotide Exchange Factor 4 (RAPGEF4). IPH was strongly related to higher expression of Biliverdin Reductase B (BLVRB) and Cyclin-Dependent Kinase Inhibitor 2A (CDKN2A), but lower levels of Nodal Modulator 1 (NOMO1). Matrix (MATX) was more nuanced, likely as it represents less defined tissue types, and was associated with Interleukin-13 (IL13), and lower levels of Nudix Hydrolase 21 (NUDT21). Several other genes were also coupled to particular tissue types by these analyses, both with and without previous associations to atherosclerosis (Supplement Table VII).

Figure 3:

Highest associations between tissue characteristics resulting from unsupervised cluster analysis. Plaque morphology is designated in shades of purple, darker indicating higher measured values and lighter indicating lower, in each of four quantitative measurands. MaxXXXXArea is the largest cross-sectional area of the indicated tissue type, MaxXXXXAreaProp is the corresponding proportional occupancy, XXXXVol is the volume counterpart, and XXXXVolProp denoting the proportional occupancy calculated on a volume basis. Transcript levels indicated in a scale with green indicating highest, red lowest, and black intermediate.

Supervised Models for Estimation of Continuous Gene Expression from Plaque Morphology

Models were then built for each transcript to estimate the continuous valued expression level based on plaque morphology. Among those that rated high for model quality, we found transcripts of two functionally different divalent cation transporters where expression clearly associated with plaque morphology, Solute Carrier Family 30 Member 1 (SLC30A1) and Solute Carrier Family 39 Member 8, encoding ZIP8 (Fig. 4A, B). High levels of CALC predicted relatively low, IPH predicted high expression of both transporters, while LRNC predicted high expression of influx (SLCA39A8) but low expression of efflux transporters (SLC30A1; Fig. 4B, C). Other examples of transcripts for which morphological assessment provided robust continuous-value estimation are listed in Supplement (Table VI).

Figure 4:

SLC30A1 as an example for which plaque morphology provided good estimation of gene expression. Continuous-value expression model performance represented by a scatter plot with multiple regression lines where each curve plots the best model fit for each of the different predictor sets considered. SLC30A1 expression prediction performance using plaque morphology was superior to clinical variables (laboratory values), and optimal when used alone rather than in combination (A; points and regression lines in colours annotated at bottom of graph). Table comparing the correlation for each of three tissue types on expression level for two transporter genes representing Zn influx (SLC39A8) and Zn efflux (SLC30A1) (B; Pearson’s r correlation shown to indicate relative direction rather than magnitude; MaxCALCArea, MaxIPHArea, and MaxLRNCArea indicates maximum cross-sectional area; Prop suffix indicates proportional occupancy of tissue relative to overall wall area). As IPH increased and CALC decreased, the expression of both transporters increased, but as LRNC increased, expression of influx transporter increased and efflux transporter decreased (C). TGF-β receptor type 2 as an example where the best model was fit combining plaque morphology and clinical factors. Continuous-value estimation of TGFBR2 expression was better when morphological assessment was utilized relative to the use of clinical variables (laboratory values) alone but was optimal when both predictor sets were combined (D; points and regression lines in colours annotated at bottom of graph). Differences observed in plaque morphology associated with increased expression of either TGFBR2 or IL1R1. Higher TGFBR2 expression was associated with high CALC burden, whereas IL1R1 was more expressed in plaques with larger LRNC (E).

Expression levels of several transcripts could also be assessed through a combination of morphological and clinical variables, which improved determination of continuous-valued expression levels, including a number of cytokines and cytokine receptors. Transforming Growth Factor-beta (TGF-β) receptor type 2 (TGFBR2) was best fit with a model combining plaque morphology and clinical factors (CCC=0.3, slope=0.8; Fig. 4D). Morphological assessment alone was superior to clinical factors alone but improved further when both variables were combined and superior to the degree of stenosis. We also observed an interesting relationship between morphology, TGFBR2 and Interleukin-1 receptor (IL1R1) expression. Whereas increased levels of TGFBR2 was predicted by lesions with high CALC and with less IPH, plaques with larger LRNC predicted higher IL1R1 expression (Fig. 4E). In support of this observation, qualitative immunohistochemical assessment of IL1R1- and TGFBR2-protein expression in CEA specimens demonstrated more IL1R1 staining in plaques with LRNC predominance whereas TGFBR2 was more abundant in lesions with high CALC (Supplement Fig. VII). Other examples of transcripts well estimated by morphological assessment alone or in combination with clinical variables are provided in Supplement (Table VI).

Models for Dichotomized Classification of Gene Expression

A second set of models were built for dichotomized classification of transcript levels above or below median expression value. For example, dichotomized expression (higher vs. lower) of microRNA 125b-1 (MIR125B1) was well classified by morphology (Fig. 5A), as well as MIR718 and MIR4536-1. Lower expression of MIR125B1 occurred in plaques with a combined burden of LRNC and IPH comparable to CALC, whereas expression was higher in smaller plaques with proportionally more CALC (Fig. 5B). In contrast, lower expression of MIR718 was found in smaller plaques with relatively high CALC burden but increased in larger plaques as CALC proportion decreased. Expression of MIR4536-1 was lower in larger plaques with increased CALC burden and decreased further in plaques with less CALC (Fig. 5C). Additional transcripts predicted as relatively high vs. low levels using morphological assessment are listed in Supplement (Table VII).

Figure 5:

MIR125B1, MIR718, and MIR4536-1 were examples where dichotomized expression level (defined as being higher or lower than the median) demonstrated robust classification accuracy. Receiver operating characteristic (ROC) curves for classifying MIR125B1 by various predictor sets, where classification performance differed based on the predictors sets used. MIR125B1 expression level was robustly estimated using plaque morphology, superior to clinical variables (laboratory values) and was optimal when used alone rather than in combination (A; ROC curves in colours annotated at bottom of graph). Volumes for plaque tissue components typifying lower (below median) expression of each of the microRNAs demonstrated different distributions of the interrogated tissue types (B). Changes in the plaque composition as expression increased showed even greater differences (C), e.g. higher CALC predict lower expression of MIR125B1 but decreased expression of MIR718 and MIR4535-1. Vertical axes in B and C relative rather than absolute scaling, and colour key is provided at right.

Estimation of Biological Processes Predicted by Plaque Morphology

The biological relevance of transcripts with expression levels predicted in dichotomized form was investigated by GSEA to expose biological processes elucidated by plaque morphology. Of the 414 robustly predicted transcripts, 237 transcripts were classified as particularly robustly predicted and hence eligible for pathway analysis as evidenced by ranking according to the product of point estimates against an objectively determined cut-off (Supplement Table VIII). Several fundamental processes related to the pathophysiology of atherosclerosis and plaque instability were found to be enriched such as SMC proliferation; ECM organization; collagen degradation; apoptosis, phospholipid and cholesterol efflux; regulation of epithelial to mesenchymal transition, and neutrophil mediated immunity (Fig. 6).

Figure 6:

Pathway analysis of the 237 transcripts meeting model quality to be described as particularly robustly predicted (for which both AUC and Kappa were high) at the cohort level as means to identify biological processes significantly determined by the robustly predicted transcripts. Identified biological processes classified by basic atherosclerotic disease mechanisms (colour codes as per the Key).

Validation of Predicted Gene Expression from Plaque Morphology

The resulting predictive models were validated by image analysis of CTAs from four patients excluded from model development, but where microarray data from CEA specimens was available for comparison. The plaque morphology heatmaps indicated different plaque characteristics: patient T1 had a lesion with a relatively large proportion of MATX, low CALC and intermediate amount of LRNC and IPH; T2 had high levels of LRNC, low MATX, and intermediate levels of CALC and IPH; while T3 and T4 were more calcified. The predicted 20 most significantly dysregulated transcripts, compared with the true expression of corresponding transcripts, demonstrated unique dominant mechanisms for each patient derived from plaque morphology (Fig. 7).

Figure 7:

Performance using locked-down models on four sequestered (unseen) test patients (T1-T4). Heatmaps for each test patient representing: plaque morphology profile; predicted expression of the top 20 most significant predicted transcripts; and true expression of corresponding transcripts obtained from microarray analysis of CEA specimens (left column). Dominant mechanisms obtained from pathway analysis by GSEA for each patient (right column). Heatmaps colour coded as in Figure 3, and biological processes in the Dominant Mechanism column as in Figure 6.

Patient T1’s profile showed dysregulation of epithelial to mesenchymal transition (adj. p=0.015), while T2’s profile resulted in 7 significant processes: collagen degradation (adj. p=0.002), ECM degradation (adj. p=0.011), regulation of membrane protein ectodomain proteolysis (adj. p=0.02), positive regulation of lipid biosynthetic process (adj. p=0.027), HDL-mediated lipid transport (adj. p=0.041), ECM organization (adj. p=0.042), and phospholipid efflux (adj. p=0.043). Patient T3 had significantly dysregulated epithelial to mesenchymal transition (adj. p=0.01). Patient T4 had two significantly dysregulated processes: regulation of SMC proliferation (adj. p=6.2e-04) and GPVI-mediated activation cascade (adj. p=0.047).

Discussion

This pilot study demonstrates the feasibility of commonly available CTA imaging to generate per-patient molecular level signatures for atherosclerotic plaque phenotyping. Decoding atherosclerotic plaque phenotype non-invasively from CTA’s with validated software determining morphological features of stable and unstable atherosclerotic lesions was supported by unsupervised clustering, multivariable predictive modelling of gene expression levels, and elucidation of patient-specific dominant mechanisms with strong statistical significance. The work provides a base for studies in larger cohorts, which may lead to future strategies for personalized therapy in the prevention of consequences to atherosclerotic plaque instability.

Models were developed for genes and non-coding RNAs most relevant to atherosclerotic plaque instability. We first identified transcripts where expression levels could be estimated from CTA analysis by matching plaque morphology against the expression of 3,478 transcripts selected based on their relevance to atherosclerotic disease.^{33, 37, 46} 414 transcripts were robustly predicted and coupled to relevant plaque features such as LRNC with biological pathways associated with inflammatory processes and ECM degradation^{49, 50} and IPH with expression of BLVRB and haemoglobin metabolism, as previously reported by our group.³³ Of the 414, 237 met the further criteria for inclusion in pathway analysis as being particularly robustly predicted. Moreover, approximately 100 of the 237 could be estimated more specifically by continuous, actual, value, thus beyond the ability to predict high or low expression only.

Some transcripts demonstrated novel associations between morphological plaque features and expression levels,⁵¹ such as transcripts encoding divalent cation transporters that may mediate effects of nitric oxide (NO),⁵² and contribute to functional regulation of macrophages and SMCs,^{30, 53} whereas others confirmed previously reported relevance in atherosclerotic plaque instability, such as CDKN2A.⁵⁴ Expression levels of several transcripts could also be determined by combining morphological and clinical variables, which improved predictive power and was superior to the degree of stenosis, a clinically used surrogate marker for stroke-risk in patients with carotid stenosis.⁵⁵ For example, in this analysis, IL1R1 expression associated with LRNC and TGFBR2 to highly calcified lesions. Previously, IL1β-mediated immune signalling through IL1R1 has been attributed a key role in atherosclerotic inflammation.⁵⁶ Inhibition of this interaction has been shown to reduce plaque progression in atherosclerotic mice⁵⁷ and improve outcome in patients with CVD.⁵⁸ In contrast, TGFβ and its receptors have been coupled to profibrotic processes and plaque stabilizing effects, which may be consistent with the association of TGBR2 to highly calcified, stable, lesions,³⁷ and here also observed adjacent to macro calcifications by immunohistochemistry. Combination of morphology and clinical variables could also predict expression levels of MMP12, previously reported by our group and others to be associated with ischemic stroke.⁵⁹

Even more transcripts could be estimated as to be expressed at relatively higher or lower levels (i.e., dichotomized). These results may partly be due to the dichotomized nature of the development cohort with lesions ranging from high to low calcification, the latter generally comprised of lipid-rich plaques. Whereas generalizability to other cohorts with more heterogeneous plaque morphology would be desired, at this stage, this is more a strength than a weakness as it allowed investigation of biological processes contributing to plaque stability in the presence and absence of macro-calcification.⁶⁰

Several miRs previously implicated in atherogenesis^{61, 62} and fibrous cap stability⁶³ were found to be well predicted. MIR125B1 has been tied to oxidative stress,⁶⁴ as well as autophagy in cancer,⁶⁵ which may be relevant as autophagy deficiency in macrophages promotes atherogenesis.⁶⁶ Transcripts identified in this analysis, such as BLVRB, IL1B, and MMP8 also emphasized the predictive potential of LRNC and IPH as morphological features with ability to determine gene expression, which could be useful in drug development.⁶⁷ In support of the clinical and biological relevance of these findings, pathway analysis of transcripts where expression levels were determined in at least dichotomized form, revealed associations to established biological processes in atherogenesis and plaque instability.

The validity of our models was finally tested in a sequestered set of patients by predicting gene expression from plaque morphology by CTA, which was compared with transcriptomic data from corresponding tissue specimens. The results demonstrated a good correlation between predicted and observed expression levels of transcripts, while pathway analysis of the most significant transcripts demonstrated unique dominant mechanisms for each individual. Notably, the analysed plaque of one patient (T2) was dominated by lipid metabolism in a manner quite different from the other patients, which suggests opportunities for patient-specific plaque phenotyping as guidance for individualized therapy.

For a QIB to have a clinical impact in the management of CVD, it is necessary to utilize non-invasive imaging routinely used in clinical practice as catheter-based imaging would not be indicated in patients with earlier stages of the disease. Whereas there is increasing evidence that manually read CTA by itself can inform about risk,⁴⁰ machine analysis is needed to categorize lesions based on specific dysregulated molecular pathways. The biological processes predicted by CTA image analysis of individual patients in this study represented relevant pathways in the pathophysiology of atherosclerosis and plaque instability, suggesting potential for utility in personalized therapy. Such analysis applied at the per-patient level could inform selection of patients for surgical or endovascular intervention, or medical therapy based in the determined dominant mechanisms, for example either targeting a plaque phenotype accessible for intensive lipid lowering or anti-inflammatory therapies.^58,68 In addition, QIBs for atherosclerotic plaque phenotyping may improve drug development by providing non-invasive markers to monitor therapeutic effects and improve efficacy of clinical trials.⁶⁹

Our study has limitations. Although the study was aimed to demonstrate the feasibility of this approach, the cohort was admittedly small, patients were selected based on carotid plaque calcification, and there was only one reader used. This should be broadened in a larger study encompassing more diverse morphological plaque features and with a larger cohort of sequestered patients used for model validation. Additionally, our study was performed in patients with end-stage atherosclerotic disease, as plaque tissue is generally not available from earlier stages of the disease. As the dynamics of atherogenesis are reflected at the molecular level, extrapolation from end-stage lesions may not apply to earlier phases of the disease. However, the patients studied represented those where treatment is required, which suggests clinical significance if the results would sustain in a larger study. Last but not least, the specific models developed did not address whether the determined associations were causal, or if so, in which direction. Whereas we acknowledge this as a limitation, these results serve as a basic framework for future studies aimed to resolve causal associations using model formalisms.

To the best of our knowledge, our “virtual transcriptomics” of atherosclerotic plaque tissue enabled by processing of conventional CTAs, is the first study to demonstrate that interrogation of plaque biology can be accomplished in individual patients across scales. The results from this pilot study will serve as a basis for further exploration of non-invasive imaging to elucidate molecular signatures in atherosclerosis and may facilitate individualized therapy in the prevention of myocardial infarction and ischemic stroke by categorizing patients according to likely response of different therapeutic choices.

Highlights.

• This pilot study demonstrates the feasibility of CTA imaging to generate per-patient molecular level signatures for atherosclerotic plaque phenotyping.

• Whereas manually read CTA by itself can inform about risk, machine analysis is needed to provide interpretation across scales.

• “Virtual transcriptomics” of atherosclerotic plaque allowed interrogation of plaque biology in individual patients across scales.

• The results will serve as a basis for exploration of non-invasive imaging to categorize patients according to likely response of therapeutic choices.

• This method can play a role in translating biological understanding through to improved clinical guidelines for optimizing care.

Abbreviations and Acronyms

AUC

Area under the receiver characteristic curve

BiKE

Biobank of Karolinska Endarterectomies

BLVRB

Biliverdin Reductase B

CAD

Coronary artery disease

CALC

Calcified tissue

CDKN2A

Cyclin-Dependent Kinase Inhibitor 2A

CEA

carotid endarterectomy

CPM

Cellular Potts models

CTA

Computed tomography angiography

CVD

Cardiovascular disease

ECM

Extracellular matrix

GSEA

Gene-set enrichment analysis

HU

Hounsfield units

IL13

Interleukin-13

IL1R1

Interleukin-1 receptor 1

IPH

Intra-plaque haemorrhage

LRNC

Lipid-rich and necrotic core

MATX

Matrix

MMP12

Matrix Metalloproteinase 12

MQ

Model quality

NASCET

North American Stenting and Carotid Endarterectomy Trial

NO

Nitric oxide

NOMO1

Nodal Modulator 1

NUDT21

Nudix Hydrolase 21

ODE

Ordinary differential equations

PDE

Partial differential equations

PRG4

Proteoglycan 4

PVAT

Perivascular adipose tissue

QIB

quantitative imaging biomarkers

RAPGEF4

Rap Guanine Nucleotide Exchange Factor 4

SLC30A1

Solute Carrier Family 30 Member 1

SLCA39A8

Solute Carrier Family 39 Member 8

SMC

Smooth muscle cell

SPDYE1

Speedy/RINGO Cell Cycle Regulator Family Member E1

TGF-β

Transforming Growth Factor-beta