Bringing doctors together

or

A free, private network for verified physicians

Leading voices

Dr. Shashank Sinha

Advanced Heart Failure · Inova Fairfax Hospital

View

Dr. Javed Butler

Advanced Heart Failure · Baylor Health Care System

View

Dr. Andrew Sauer

Advanced Heart Failure · Saint Luke's

View

Dr. Anu Lala

Advanced Heart Failure · Mount Sinai

View

Dr. Robert Mentz

Advanced Heart Failure · Duke University

View

Dr. Sunny Varshney

Advanced Heart Failure · Stanford

View

Dr. Stephanie Hsiao

Advanced Heart Failure · Stanford

View

Dr. Michelle Kittleson

Advanced Heart Failure · Cedars-Sinai

View

Dr. Adrian Hernandez

Advanced Heart Failure · Duke

View

Dr. Dorairaj Prabhakaran

General Cardiology

View
  • Discover your community

  • Stay current

  • A profile that represents you

  • Find your people

About RoonPrivacyTerms

    Overview

    Proteogenomic Analysis of Coronary Artery Calcification in Human Populations.

    Apr 2, 2026DOI: 10.1161/atvbaha.125.324171Journal ArticleAHA

    Written by: Bassim El-SabawiXiaoning HuangPhillip LinMohammad Yaser AnwarMichael BettiNamju KimAndrew S PerryB Lakshitha A PereraPriya GajjarLaura A ColangeloKaushik AmancherlaQuanhu ShengShilin ZhaoLindsay K StolzeEric Farber-EgerJoshua M LandmanPatricia E MillerGabrielle Y LiuSuman DasQuinn S WellsJames G TerryDonald Lloyd-JonesSaumya DasSadiya S KhanKari E NorthJennifer BelowMatthew NayorRavi KalhanJohn Jeffrey CarrEric R GamazonRavi V Shah

    Find these authors on Roon

    KA
    Dr. Kaushik Amancherla
    Advanced Heart Failure · VUMC

    INTRODUCTION

    Atherosclerotic cardiovascular disease (CVD) is the most common cause of cardiovascular demise worldwide1. CVD has origins decades before the occurrence of key clinical events, including sudden cardiac death and myocardial infarction2. Despite ongoing, broad pharmacologic and behavioral efforts to mitigate risk, progressive worsening in dysmetabolic phenotypes has promoted an increase in early CVD, with nearly 50% of CVD mortality under age 65 in some reports3,4. Accordingly, defining underlying targetable CVD mechanisms at an early subclinical stage is a major goal in modern precision cardiovascular prevention. Coronary artery calcification (CAC)—a touchstone phenotype of early CVD susceptibility with therapeutic implication5—represents a culmination of interconnected molecular events within endothelium and the vascular wall involving local and systemic inflammation, immune activation, and metabolic-oxidative stress6. Molecular studies of early coronary disease progression in humans have focused on relating accessible genetic or metabolic states (via metabolites, proteins, peripheral blood transcriptomes)7–9 to CAC or clinical CVD. Nevertheless, examining underlying determinants of coronary artery disease—including in a potentially tissue-specific fashion, limited due to inaccessibility of coronary tissue6—is critical to identify and interrupt early, intervenable changes in human CVD.Here, we integrate proteomics and genomic techniques toward two major goals: (1) to identify circulating proteins related to presence, extent, and progression of CAC; (2) to map these targets into human coronary artery tissue through human genetic and transcriptional techniques. For the former (and primary) goal, we identified circulating proteins associated with both prevalent and 10-year incident CAC phenotypes in ≈3,000 individuals in the Coronary Artery Risk Development in Young Adults (CARDIA) study, with validation in the Framingham Heart Study (FHS; both using extant data). For our secondary goal, we examined whether targets identified at the protein level have evidence for potential causal roles in CAC and downstream CVD. To accomplish this, we employed two parallel approaches: (1) genetic approaches targeting the circulating proteomic correlates of CAC (via Mendelian randomization [MR] and proteome-wide genetic association studies [PWAS]) and (2) genetic approaches targeting gene expression in human coronaries (transcriptome-wide association study [TWAS] of CAC in >35,000 individuals based on genetic models to predict gene expression in human coronary arteries from the GTEx consortium). These approaches utilize human genetic variants (or models based on multiple variants) associated with circulating proteins (MR and PWAS) or coronary artery transcripts (TWAS) to perform an association study with the phenotype or clinical outcome of interest; given that the genetic variants are assigned in meiosis, these results are less susceptible to confounding than proteomic associations. As with our proteomic findings, we present our PWAS and TWAS results as a scientific resource for further discovery. While transcriptional events in the coronary wall are not necessarily reflected in the circulating proteome, we explored potential overlap across the TWAS and proteomic associations of CAC in human coronaries arteries with and without atherosclerosis at a single-cell level to identify potential concordances that may further parse targets for future mechanistic validation. Ultimately, this work reports a proteogenomic analysis of liability to coronary calcification—an important harbinger of clinical vascular events—identifying both circulating and tissue-relevant determinants of CAC for therapeutic and biomarker validation in future studies.

    MATERIALS AND METHODS

    All data and materials have been made publicly available at the CARDIA Coordinating Center (https://sites.uab.edu/cardia) and can be accessed via e-mail to coc@uab.edu. No new code was developed as part of this study. Data from the Framingham Heart Study (FHS) is available at dbGaP (dbGaP identifier pht006013) or via contact with the FHS coordinating center (www.framinghamheartstudy.org). UK Biobank data (accessed under application number 94960) is available at UK Bio bank Research Access Portal. Code used for analysis in this study are available on Zenodo (for TWAS models: https://doi.org/10.5281/zenodo.3842289) and on the Github repository for this project (https://github.com/gamazonlab/CACOmics).
    A total of 2,971 participants with circulating proteomics and CAC scores by computed tomography at the Year 25 visit were included10. Clinical-demographic data was collected by standardized assessment as described11,12, and computed tomography for CAC was performed as previously noted via standardized protocols13,14. In addition, 793 participants had additional computed tomographic CAC data collected at Year 35 in CARDIA for analysis of incident CAC development by Year 35 (in those individuals without CAC at Year 25). Study participants provided written informed consent, and the study was approved by the respective clinical site Institutional Review Board.
    We studied 573 individuals with extant proteomics and computed tomographic CAC measurements from the FHS Generation 2 (“Offspring”) cohort to assess for replication of our findings in CARDIA15. CAC was performed a median of 17.4 years after collection of the blood sample in FHS (that was eventually used for proteomics). Methods for clinical and calcification traits have been reported16,17. All FHS participants gave written informed consent, and all study protocols received approval from the Institutional Review Board at Boston University Medical Center.
    In CARDIA, we used the SomaScan proteomics platform (aptamer-based technology) to quantify 7,228 aptamers representing 6,402 unique proteins in CARDIA (Somalogic, Boulder, CO), as reported in previous work10. We excluded 71 aptamers (41 unique proteins) with a coefficient of variation exceeding 20%. After log2 transformation and standardization of aptamer levels, aptamer levels were winsorized to limit the influence of extreme values. Specifically, standardized aptamer values exceeding 5 or −5 (i.e., ±5 standard deviations from the mean, calculated per aptamer) were recoded to 5 or −5, respectively.
    To explore generalizability of our findings, we studied population-level proteomics in FHS. In FHS, protein profiling was completed in two batches. In batch one, 1,129 aptamers were profiled in 821 individuals, and batch 2 included an expanded panel of 1,372 aptamers, which was assayed in 1092 participants. Aptamer levels of 1,128 aptamers common to both batches were log-transformed, standardized within batch, then pooled and normalized to a mean value of 0 and SD unit of 1 using the Blom rank-based inverse-normal method. These values were then regressed on assay plate ID to account for plate effects, and the standardized residuals were used for regression. Methods for handling proteomic data (batch pooling, transformations, and rank normalization) have been reported18; we have reproduced them with minimal change to maximize reproducibility. Matching of CARDIA to FHS aptamers was performed by SomaScan seqID (1,035 overlapping aptamers representing 1,010 unique proteins by UniProt identifier).
    In CARDIA, protein aptamers were log transformed, centered, standardized to unit variance, and winsorized to 5 standard deviations for regression. Given absence of known datasets with harmonized proteomic platform and concurrent measures of CAC, we randomly split our CARDIA sample into a derivation (70%) and validation subsample (30%). We estimated models in the discovery subsample for presence/absence of CAC (as a binary variable in logistic models) and its extent (continuous in linear models; modeled as ln[CAC+1] to account for zero CAC values). While ln(CAC+1) transformation does not fully normalize the distribution, it is a commonly used approach in prior studies to accommodate the skewed nature of CAC data for statistical modeling19–21. Each model contained one aptamer at a time, adjusted for age, sex, and race (Black or White), with a 5% false discovery rate (Benjamini-Hochberg FDR) applied across models to control for multiplicity. Aptamers that were significant at a 5% FDR threshold were passed to identical models in our validation subsample. Aptamers that were significant at 5% FDR for both presence of CAC and its extent in both derivation and validation subsamples, were analyzed in similar logistic regression models (adjusted for age, sex, and race) for incident CAC among subjects who had CAC scores of 0 at Year 25 and available computed tomography data at Year 35. Aptamers in the incident CAC models were considered significant using a threshold of 5% FDR. We included all aptamers regardless of duplication as a transparent scientific resource for replication for the community. Proteomic associations with prevalent CAC (Year 25) were performed using SAS. Proteomic associations with incident CAC (Year 35) and data visualization of all proteomic associations were performed using R 4.3.1 (packages ggplot23.4.3, ComplexHeatmap2.18, circlize_0.4.16).
    Replication studies in FHS were performed in a similar fashion (logistic for presence or absence of CAC; continuous linear for extent as standardized ln[CAC+1]), with each protein in separate models adjusted for age, sex, and race (white vs. non-white) with 5% FDR used to control type 1 error. We assessed concordance of CARDIA and FHS effect size via Pearson correlation.
    We identified protein quantitative trait loci (pQTLs, defined here as genetic instruments associated with circulating protein levels at genome-wide significance) that were measured with 4,907 aptamers in 35,559 Icelanders (summary statistics obtained from prior study22 at decode.com/summarydata/; quality control reported previously). We conservatively selected pQTLs exceeding a genomic threshold (P<5×10−8) for the target proteins, leveraging both cis-pQTLs (+2 Mbp on both directions of the encoding locus) and trans-pQTL (when available). While trans pQTLs are more likely to regulate multiple targets pleiotropically, for the purpose of Mendelian Randomization (MR), inclusion of trans pQTLs may improve protein-phenotype associations as they can increase MR power analysis23. This subset was subject to linkage disequilibrium (LD) clumping (R2<0.001, 10kbp window, 1000 Genome European reference). We restricted pQTLs to those exceeding an F statistic over 10 to minimize downstream weak-instrument bias. Of 131 unique proteins (by UniProt identifier) that were associated with presence and extent of CAC at a 5% FDR in the derivation and validation samples in CARDIA, we identified at least one pQTL for 100 proteins. For disease states, we selected three of the largest GWASs capturing CAC-relevant clinical disease states for assessment in Mendelian randomization: (1) coronary calcification (N=36,720 Europeans; GWAS catalog accession GCST90278455)21; (2) coronary artery disease (CAD; N=1,165,690 Europeans; GWAS catalog accession GCST90132314)24; (3) myocardial infarction (N=639,221 Europeans; GWAS catalog accession GCST011365)25.
    Our two-sample MR framework tested whether proteins implicated in our initial epidemiologic and tissue evaluation had evidence of potential causality by human genetic approaches to CAC and relevant outcomes. We used pQTLs (as instrumental variables [IVs] for genetically determined protein levels) as exposures and our three CAC-relevant clinical disease states as outcomes. Quality control was performed using TwoSampleMR R 0.6.21 package2 with correction for strand mismatches via the harmonise_data() function). All primary variants present in both the outcome and predictor datasets were considered for the analysis; proxy variants were not included. Compatibility of genome builds between the summary statistics used for the outcome and predictors (pQTLs) was confirmed prior to analysis. We primarily tested potential causal associations with inverse variance weighted robust-penalized method, using MendelianRandomization 0.10.0 R package3. We used ‘default effect’ model for association testing, or ‘random effect’ model when IVs for the protein demonstrated significant heterogeneity in effect size (Q-het P<0.05). Given that the candidate targets that we passed on to MR analyses already had evidence from our proteomic association studies, we considered the protein as potentially causally associated with the CAC-relevant outcome using a nominal P<0.05 threshold. In addition, we used median penalized (to limit IV-related bias) and Egger-penalized robust methods (for pleiotropy) as sensitivity analyses to enhance our confidence in causal interpretations, consistent with STROBE-MR guideline for causal inferencing using observational data26. Our MR analyses were performed in R 4.4.0.
    As a secondary approach, we performed proteome-wide association studies (PWAS) of CAC, using genetic models of circulating proteome developed in a large-scale proteo-genomic study (N = 7,213 , European ancestry; 4,657 aptamers corresponding to 4,435 unique genes)27. We included only those proteins that attain nominal significance during model training (P<0.05, cross-validation r=0.10, 1,340 unique proteins). This resource does not overlap with the UK Biobank (N = 361,194), allowing us to estimate the association between the genetically determined circulating protein expression and related traits in an independent phenomic resource. We used GWAS summary statistics from a linear model of a phenotype as a function of the first 20 genotype-based PCs, sex, age, age2, sexage, sexage2, and the protein under test to estimate the effect size of the protein on the two phenotypes in the UK Biobank: coronary atherosclerosis (code I9CORATHER) and myocardial infarction (code I9MI). We tested proteins associated with presence and extent of CAC in CARDIA (derivation P<0.05) with unique gene symbols in PWAS.
    We developed gene expression models in human coronary artery samples from 268 unrelated individuals in the Genotype-Expression Tissue (GTEx) database. This approach uses extant (but not previously widely used) bulk RNA-seq transcriptomic data in human coronary arteries alongside whole genome sequencing data in GTEx to develop genetic models of transcript expression that can be applied more generally to allow genetic inference in large GWAS or other genomic resources. We used our joint-tissue inference (JTI) transcriptome-wide association study (TWAS) methodology, which has been reported and validated28. JTI is an approach to test the effect of genetically determined gene expression on phenotype. JTI consists of a set of gene expression models using cis genetic variants as features. The training of models is improved by use of shared regulatory architecture between tissues through joint-tissue analysis of a unique multi-tissue (49 tissues) transcriptome resource (GTEx). The approach also integrates a causal inference methodology on the effect of gene expression on the phenotype under study. The JTI models can be applied to GWAS data to identify gene-phenotype associations. The training data used to develop the JTI models was GTEx (v8). We trained prediction models of normalized gene expression using sex, platform, and 5 genotype-based principal components as adjustments. We also performed phenome-wide association studies (PheWAS) by identifying the associations of the genes with the heritable phenome (heritability h2 P < 0.05) as captured in the UK Biobank.
    We leveraged single-cell RNA sequencing (scRNA-seq) data derived from 13 human coronary arteries (8 with atherosclerotic lesions and the remainder lesion-free controls) involving 56,183 cells29. We analyzed the data using the Seurat R 5.2.1 package30. Doublet removal was performed using scDblFinder 1.18.031. scDblFinder used 3 iterations and only consensus cell IDs from all 3 runs were retained as (consensus) doublets (following Mosquera et al.29). Cell barcodes identified as doublets were then filtered out from raw counts matrix. As ambient RNA contamination can negatively affect gene expression profiling, we performed correction via DecontX using default parameters32. We applied raw count normalization through a regularized negative binomial regression approach as implemented in SCTransform 0.4.133 (with argument vst.flavor = ”v2”). To avoid confounding by cell cycle state, we adjusted for cell cycle variance. For integration of processed sequencing libraries across potential technical differences and dataset-specific conditions, we used Harmony 1.2.334. We applied dimensionality reduction using principal component analysis (PCA) on the normalized counts, leveraging the first 30 principal components (PCs) for clustering. We invoked the functions RunUMAP (using 30 neighboring points for local approximation of manifold structure) and FindNeighbors (using k=20 for the k-nearest neighbor algorithm) in Seurat, setting reduction to ‘harmony’ to use the Harmony embeddings. Cell type annotation leveraged Seurat’s reference-based transfer learning applied to the Tabula Sapiens (specifically vasculature data in 42,650 cells available at https://cellxgene.cziscience.com/e/a2d4d33e-4c62-4361-b80a-9be53d2e50e8.cxg/)35. For Tabula Sapiens, we applied the same SCTransform normalization workflow and UMAP dimensionality reduction as used for the samples, ensuring consistency with the cell annotations. We examined differential expression between the samples with atherosclerotic lesions and the lesion-free samples using pseudobulking-based DESeq236. We used 5% false discovery rate to control type 1 error.

    RESULTS

    Our overall study scheme is shown in Graphic Abstract. Clinical characteristics of the study populations (Coronary Artery Risk Development in Young Adults; CARDIA; N=2,971; Framingham Heart Study; FHS; N=573) are in Table 1. CARDIA participants at Year 25 had a mean age of 50.2±3.6 years, with approximately even distribution by sex and self-identified race (56% women, 46% Black) and prevalent cardiometabolic risk (mean systolic blood pressure 119±15 mmHg, 8% diabetes, mean BMI 30.3±7.1 kg/m2) and CAC (≈29%). A subgroup of participants with baseline proteomics and without prevalent CAC at Year 25 had computed tomography performed ≈10 years later (Year 35 study visit in CARDIA; N=793), with an incident CAC rate of ≈31%. FHS participants were similar in age and sex, with a low prevalence of clinical cardiovascular disease (≈4%) at the time of proteomics, with a high rate of CAC (≈69%; when measured ≈20 years after blood sample was collected for proteomics).
    We identified 136 aptamers (representing 131 unique proteins) associated with presence and extent of CAC across discovery and validation sets within CARDIA (Figure 1A for example volcano plot for CAC extent; presence of CAC, 186 aptamers, 174 unique proteins; extent of CAC, 201 aptamers, 193 unique proteins; regression summaries in Table S1). Of note, proteins with multiple aptamers significant against either extent or presence of CAC had concordant directionality across aptamers. Identified proteins exhibited strong biological plausibility across known mechanisms of vascular remodeling (Table 2), including fibrosis and inflammatory mechanisms (GDF-1537, CDCP138, GSN39, THBS240, chemokines, RNAS641), oxidative lipid metabolism (CILP242), extracellular matrix remodeling and signaling (MMP-7, MMP-1243, TIMP-1, integrins44), calcification (Notch 145, ARHGAP3646), and general metabolism (GIP47). Importantly, a majority of identified protein associations (though biologically plausible) had not been previously widely reported in human coronary disease or calcification, including mechanisms of protein catabolism (UBE2G248), macrophage efferocytosis and lipid metabolism (TREM249), endothelial cell states (EGFR50), and extracellular matrix metabolism (PCOLCE).
    To address epidemiologic limitations on cross-sectional associations (reverse causation), we quantified the relation of 136 aptamers in our prevalent CAC proteome with incident development of any CAC ≈10 years later in a subsample of CARDIA that was CAC-free at time of proteomics (N = 793; Figure 1B, Table S2). Among aptamers related to the prevalence and extent of CAC at baseline, we observed strong consistency in directionality of effect (129/136) with incident CAC at 10 years (with 59 aptamers significant for incident CAC after FDR correction). In addition, we examined associations with CAC measured ≈20 years after the blood sample (on which proteomics was performed) in FHS, observing broad consistency in effect directionality and size for proteins with significant relationships for CAC in CARDIA (Figure S1; full regression results in Table S3). This includes proteins with complete agreement in effect size and significant relationships in FHS, including MMP-12 (higher levels previously associated with more coronary atherosclerosis51,52) and Contactin-1 (higher levels previously associated with absence of coronary atherosclerosis51,52).
    Next, we performed MR studies to increase confidence in the targetability of select mediators and their causal relevance to human CVD. For MR, we used established protein quantitative trait loci (pQTL) for proteins associated with presence and extent of CAC in CARDIA (100 included that satisfied quality control thresholds; see Methods). We observed broad consistency between MR-based genetic estimates of effect on three major disease conditions (CAC, coronary artery disease, and myocardial infarction; spanning ≈1.8 million individuals) and proteomic-phenotype associations from CARDIA (Figure 1C, full results in Table S4), several of which have not been widely reported in coronary vascular calcification in humans. While a full exposition of findings is out of scope for this work, several findings merit mention. Proteins with both proteomic and genetic evidence in favor of increased calcification or clinical risk (at P<0.05 for at least one MR method) included mediators involved in central mechanisms of metabolism and inflammation, including GABARAPL153 (autophagy), S100A9/calgranulin B (inflammatory signaling/atherosclerosis in murine models54), VEGFA55,56 (angiogenesis and lipid handling), C1QTNF157 (macrophage activation). Several genes with evidence of protection against endpoints at a proteomic and genomic level had not been broadly implicated in calcification in humans, including proteins implicated in central nervous system development and physiology (SLITRK158,59: noradrenergic system development; IGLON560; NTRK3), extracellular matrix metabolism (HS6ST3, ADAM23) and fibroblast development (FGFR1).
    In a parallel approach, we estimated the association of genetically determined protein abundance with atherosclerotic disease phenotypes in 361,194 UK Biobank participants through proteome-wide genetic association studies (PWAS). Our studies prioritized those proteins associated with CAC in CARDIA proteomics (in derivation P<0.05) with suitable PWAS models (see Methods). At a 5% Bonferroni threshold (across the tested proteins), we observed four proteins with significant evidence for involvement of genetically determined protein abundance in atherosclerosis (PCSK9, APOC1, HTRA1, and FGFR1; Figure 1D) and two for myocardial infarction (PCSK9, INHBC; Figure 1E; see Table S5). Notably, the directly measured CAC-associated proteins in CARDIA were substantially enriched for genetically determined CAC associations by PWAS (Figure 1F), with many of the targets among the most significant associations from PWAS. For example, circulating PCSK9 expression is linked to greater severity and vulnerability of coronary plaque61, with successful PCSK9 interruption linked to lower vascular event rates62,63. APOC1 (lipoprotein metabolism) has been implicated in human GWAS or functional studies of atherosclerosis64,65. INHBC (implicated in adipocyte metabolism) appears to link cardiometabolic health to coronary risk, with higher levels related to higher coronary risk66.
    Circulating proteome-wide association does not resolve target tissue-specific mechanisms of disease pathogenesis, and circulating levels of proteins related to CAC may reflect systemic factors, not necessarily those operative within the vessel wall. TWAS in a mechanistically relevant tissue may address this limitation, estimating genetically determined transcript expression within a trait-related tissue (in this case, human coronary artery) to use in conjunction with cognate phenotype GWAS data (in this case, CAC). We constructed genetic models for gene expression in 268 human coronary arteries (see Methods), mapping these models to the largest extant GWAS of CAC21 to identify gene-level associations with CAC. Coronary artery expression of genes associated with CAC in TWAS (Figure 2A; full results in Table S6) included known (PHACTR167, MRAS67, MORF4L1-ADAMTS768) and not previously widely implicated targets (GIGYF1, DMPK, RPL9). PHACTR1 (phosphatase and actin regulator-1) and MRAS demonstrated strong association with CAC, consistent with pathogenic roles in endothelial dysfunction, CAC, or inflammatory cell function67,69,70. Interestingly, MORF4L1 and ADAMTS7 are adjacent at 15q25.1 with the lead SNP in an intronic region of MORF4L168, leading to difficulty in resolving the actual causal disease gene. While only ADAMTS7 has wide support in models of atherosclerosis68, our TWAS results (based on genetically determined RNA expression in coronary artery) along with MR to account for LD contamination implicated both genes in CAC pathogenesis. Other genes identified in coronary TWAS of CAC displayed phenotypic association with anatomically or functionally adjacent processes, though not directly with atherosclerosis itself (GIGYF1: type 2 diabetes71; DMPK: cardiac conduction72; RPL9: ribosomal metabolism73).
    Over-representation analysis of top TWAS-implicated genes (Reactome; P<0.001, 36 genes, Table S6) through WebGestalt74 yielded nominally significant pathways (P<0.05; Figure S2). Several displayed an enrichment ratio of >80 (calculated as the number of input genes in the pathway divided by the number expected based on the background protein-coding genes), including “HDL clearance” (notable given HDL’s protective role against calcification75), “Pre-NOTCH Processing in the Endoplasmic Reticulum” (dysfunction can shift vascular cells towards the calcification process45), and “Methionine salvage pathway” (linked to coronary calcification via epigenetics76).
    While proteins in circulation related to CAC may not reflect events in the coronary wall, we finally explored whether any overlap between the coronary-artery specific gene targets from TWAS and our CAC-proteome associations would implicate genes of interest that may have roles in both coronary tissue and circulation. We identified 21 genes with support from circulating proteomic association with CAC in CARDIA as well as in human coronary artery TWAS of CAC (at a P<0.05 threshold; Figure 2B; Table S7). Importantly, several of these genes have previously been implicated in vascular, metabolic or inflammatory mechanisms of atherosclerosis, including PNPO (candidate gene for hypertension77), SPARCL1 (maintenance of vessel wall structure, lower in activated endothelial cells78), HES1 (responsible for cardiac arterial development79), S100A9 (vascular inflammation80) and ASGR1 (inhibition decreases hypercholesterolemia81), and other known mediators of vascular dysfunction or plaque homeostasis (NOTCH382, TNFSF1283, S100A1284). In addition, while not previously directly implicated in CVD, GPC6, RPP25, and HS6ST3 are mechanistically adjacent: GPC6 blocks canonical Wnt signaling85, a pathway central to vascular calcification86,87, and RPP25 (a part of ribonuclease P) has been implicated in genetic studies of cardiometabolic disease88 and in scleroderma (a disease marked by multi-organ calcification and fibrosis). Genetic variants in HS6ST3—a gene whose product has been implicated in protein-extracellular matrix interactions—have been linked to CAC in some settings89, though implications in human atherosclerosis have not been investigated. Finally, these 21 genes were expressed across broad cell types in human coronary arteries29, with evidence of differential expression in key cell types between atherosclerotic and non-atherosclerotic coronary tissues (e.g., S100A9 higher in macrophages in atherosclerotic tissue, with several lines of evidence across our entire study; Figure 2C–D). (Results for pseudobulk differential expression are in Figure S3.)
    Examination of our quantile-quantile plots suggested that 8 genes (NOTCH3, SPINK2, S100A12, RPP25, OAF, HS6ST3, TNFSF12, GPC6) exhibited the greatest divergence from expectation from our overall 21 genes (Figure 2B). These 8 genes were therefore prioritized for final MR approaches and PheWAS to assess whether these targets would have genetic support for a role in coronary calcification (Figure S4). We performed inference via MR with gene expression as “exposure” and CAC as “outcome”. For causal inference, we used five MR approaches, each relying on strong assumptions on the underlying ground truth: MR-Egger90, median-based91, maximum likelihood92, inverse-variance weighted93, and MR-JTI28 (Figure S4A). For 5 of these genes (SPINK2, RPP25, OAF, HS6ST3, GPC6), the direction of effect observed in MR analyses was consistent with the direction of the circulating proteomic association with CAC. In phenome-wide association studies (PheWAS) across the heritable phenome in UK Biobank (heritability h2 P < 0.05) using genetically determined expression in coronary artery, many of the 8 prioritized genes implicated vascular, inflammatory, and obesity-related phenotypes (Figure S4B).

    DISCUSSION

    Here, we unite population-wide circulating proteomics, non-invasive coronary imaging, and circulating and coronary artery tissue-specific genomics to present a framework for prioritizing targets for CVD susceptibility and downstream mechanistic work. Our studies within a large population sample (CARDIA) indexed a broad human proteome to prevalent and incident CAC—an independent marker of subclinical disease and incident clinical CAD and death prior to age 60 years11—implicating pathogenic mechanisms across oxidative stress and lipid metabolism, extracellular matrix remodeling, immune cell function, coagulation, and inflammation7,94,95. Given our rationale that coronary artery-specific contextualization may focus discovery for CVD, we next performed the largest human coronary artery TWAS of CAC, mapping bulk transcriptomics in 268 human coronary arteries to the largest CAC GWAS, demonstrating enrichment for genes implicated by proteomic associations across similar pathways of metabolic-inflammatory activation expressed in key cell types relevant to coronary disease progression. These two results—PWAS and TWAS of human CAC—have not been previously reported to our knowledge and are provided here as a discovery resource. Finally, while we recognize that genes implicated in CAC from coronary artery tissue may be distinct from circulating proteomic targets, we explored targets with evidence across both modalities using MR and single cell transcriptomics, identifying several targets with multi-level support or prior biological curation relevant to vascular disease. Collectively, these results furnish a framework to decipher disease-relevant targets through integration of human genetic approaches with multi-omics.
    In addition to well-known mechanisms of inflammation and fibrosis (Table 2), proteomics identified directionally consistent targets with strong evidence in model systems (less in humans), including TREM2 (attenuates macrophage uptake of oxidized lipids, limiting progression of atherosclerosis in mice49), Notch1 (loss potentiates RUNX2-mediated calcification45), and ARHGAP36 (overexpression associated with connective tissue to bone transformation46), among others (Table 2). Importantly, we observed a strong concordance between effect estimates for prevalent and 10-year incident CAC in CARDIA with strongest agreement for both known (e.g., MMP-12, GDF-15) and emerging mediators of vascular risk. Genetics (MR) suggested directional consistency in genomic-proteomic effects for many targets, many of which have not been extensively implicated in human coronary disease by GWAS studies but have relevant metabolic-vascular mechanisms (e.g., HS6ST396, PTPRD97, C1QTNF157, DNJB998). PWAS approaches (genomic predictions of the proteome) were also supportive, identifying one of the strongest therapeutic targets in coronary disease (PCSK9), now with support in primary prevention for vascular events (in line with our genetic findings)63.
    A second new finding of our approach is deploying a transcriptome-wide association study based on human coronary artery transcription (from GTEx) in the largest GWAS of CAC. This TWAS approach has recently been used to identify targets essential to cardiometabolic therapeutics99, though its application to coronary arteries has—to our knowledge—not been reported. The current report represents the largest TWAS of CAC in human coronary arteries to date. Human coronary TWAS associations with CAC included known (PHACTR167, MRAS67, MORF4L1-ADAMTS768) and not widely implicated targets (GIGYF1, DMPK, RPL9). In addition, pathway analysis of TWAS hits identified pathways putatively involved in calcification, including HDL clearance and NOTCH. Given the high concordance of therapeutic targets based on genetic supportive evidence100, we believe this data may provide an important starting point for further downstream mechanistic studies to inform potential therapeutic targets.
    In addition to an unbiased TWAS approach, we explored targets implicated both by proteomic associations with CAC as well as in the CAC TWAS. This approach has some precedent in recent studies linking the human proteome to tissue-specific transcriptional states101–106, though we certainly recognize that events in the coronary wall may not necessarily be reflected in the circulating proteome (and vice versa). The overlap between the coronary artery TWAS of CAC and the circulating proteomics of CAC (from CARDIA) highlighted 21 genes, some with known implications in vascular, inflammatory or metabolic mechanisms in model systems (NOTCH382, TNFSF1283, among others) and some not previously studied. Certainly, lack of concordance between circulating protein abundance and coronary tissue expression may limit discovery, and proteins in circulation may reflect systemic processes that influence CAC development. Nevertheless, convergent support from previous studies, proteomics, and coronary tissue TWAS highlight these as biologically plausible candidates for further mechanistic and therapeutic evaluation. This fundamental observation highlights the importance of studying both the peripheral circulation (where targets like PCSK9 may be found in proteogenomic approaches) as well as the coronary tissue itself as parallel, orthogonal approaches to coronary disease discovery.
    Our study has several important limitations. Importantly, our analyses proceeded in parallel (proteomic and proteogenomic studies in “Goal 1” and genetic-transcriptomic studies in “Goal 2”; Figure 1) in recognition that CAC susceptibility is driven by systemic, peripheral factors as well as coronary-specific factors. Our discovery of new proteins not quantified on this platform (Somalogic) are limited by design and can be addressed as proteomic platforms advance in coverage. Further, coronary arteries in our genomics approaches may not necessarily reflect the range of coronary phenotypes relevant to disease discovery, and we were not able to ascertain the exact histologic classification of the coronary arteries used in TWAS from GTEx. While differences in directionality of effect between tissue MR approaches and protein association are possible, our goal was to identify potential targets for downstream interrogation. While our circulating proteomic studies utilized standard type 1 error control methods (false discovery rate), we utilized more lenient thresholds in our genomics analyses, given that these were meant to be multiple lines of supportive evidence beyond circulating proteome. Also, additional efforts across proteogenomic approaches that involve multiple ancestries is critical to understanding the basis for coronary disease susceptibility across populations, given then European-ancestry predominant cohorts used here. Finally, mechanistic studies are needed to provide biological validation and final prioritization of targets for therapeutic interruption. Nevertheless, multi-level discovery in tissues and populations provide a focus for calcification mechanisms and biology for wider studies and therapeutic approaches.
    In conclusion, we present an approach to address coronary artery calcification susceptibility across the circulating proteome, coronary artery-specific transcription and TWAS, implicating genes both known and not previously reported in CAC pathogenesis. Targets identified by our approach enjoyed multi-level support across proteogenomic approaches, several with previously reported biological implications in CVD or CVD-adjacent mechanisms. This approach represents a multi-dimensional approach to human coronary atherosclerosis, incorporating both circulating and tissue-derived marks of disease liability to identify multi-level determinants of coronary disease for future mechanistic studies.
    KADr. Kaushik Amancherla

    Dr. Kaushik Amancherla

    Cardiology · Advanced Heart Failure & Transplantation
    Vanderbilt University Medical Center
    kaushik