Integrative statistical genetics prioritizes candidate ASCVD susceptibility genes across tissues

Abstract

Background

Atherosclerotic cardiovascular disease (ASCVD) remains a leading cause of mortality worldwide, yet its genetic architecture and underlying mechanistic hypotheses have not been fully elucidated.

Methods

We applied genomic structural equation modeling (SEM) to five ASCVD-related phenotypes with a total combined sample size of approximately 3.8 million individuals across the five component GWAS. Leveraging these genetic insights, we performed a cross-tissue transcriptome-wide association study (TWAS) using the UTMOST framework by integrating ASCVD genetic data with gene expression from 49 tissues (GTEx v8). Complementary analyses included single-tissue TWAS (FUSION) and gene-level analysis (MAGMA). By triangulating evidence from these multiple convergent statistical approaches, we identified robust candidate genes, which were further prioritized using Mendelian randomization, Bayesian colocalization, phenome-wide association studies (PheWAS), tissue/cell-type enrichment analyses, and AlphaFold-based structural predictions integrated with chemical interaction profiling.

Results

We identified 14 genes robustly associated with ASCVD susceptibility, including four genes—ARVCF, GFPT1, NFU1, and USP39—some of which have been previously implicated in coronary artery disease genetics (ARVCF) or are newly prioritized in the composite ASCVD phenotype (USP39), exhibiting broad tissue expression patterns. Mendelian randomization and colocalization provided supportive evidence consistent with potential causal relationships, with pronounced tissue specificity (e.g., ARVCF in artery aorta, GFPT1 in heart atrial appendage, NFU1 in adipose subcutaneous). Cell-type enrichment highlighted erythroid progenitor cells and immune populations. PheWAS revealed potential horizontal pleiotropy, while network and pathway analyses implicated these genes in cell adhesion, amino sugar metabolism, iron-sulfur cluster assembly, and RNA splicing. Structural predictions confirmed protein model reliability, and chemical interaction profiling linked these genes to cardiovascular and metabolic disease pathways.

Conclusions

Our integrative statistical genetics approach advances the functional understanding of ASCVD genetics, implicates four genes in disease pathogenesis with tissue-specific mechanisms, and highlights promising candidate genes for future investigation.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s10020-026-01554-w.

Introduction

Atherosclerotic cardiovascular disease (ASCVD)-a spectrum of conditions encompassing coronary heart disease (CHD), stroke, transient ischemic attack (TIA), abdominal aortic aneurysm (AAA), and peripheral arterial disease (PAD)-remains a leading cause of morbidity and mortality worldwide, accounting for approximately 18.6 million deaths annually, or nearly one-quarter of all-cause mortality globally (Tsao et al. 2023, Fu et al. 2025, Cao et al. 2021). Characterized by progressive lipid deposition, chronic inflammatory activation, and endothelial dysfunction, ASCVD arises from intricate interactions among genetic predisposition, environmental exposures, and lifestyle factors (Global 2025, Yvan-Charvet et al. 2025, Zeng et al. 2021). With the accelerating pace of global population aging, the incidence and public health burden of ASCVD have risen steeply, posing formidable challenges to medical research, health systems, and socioeconomic development (Global , Nedkoff et al. 2023). Despite considerable therapeutic advances in lipid-lowering therapy, anti-inflammatory strategies, and vascular regeneration, the precise genetic and molecular mechanisms governing ASCVD pathogenesis remain incompletely defined (Zhang et al. 2023). Although dysregulated lipid metabolism, chronic inflammation, and endothelial dysfunction are established core drivers, they do not fully account for the substantial interindividual variability in disease susceptibility or clinical outcomes (Nayor et al. 2021, Mitsis et al. 2025, Ma et al. 2021, Zhang et al. 2023).

To capture the shared genetic architecture across these clinically related but heterogeneous conditions, we applied genomic structural equation modeling (SEM) to five ASCVD-related traits (Grotzinger et al. 2019). A common-factor model showed excellent fit (CFI = 0.97); alternative models did not substantially improve fit (see Methods and Results). We acknowledge potential heterogeneity—the latent factor may be disproportionately influenced by CHD and PAD due to larger GWAS sample sizes—but this integrative strategy mitigates single-biomarker confounding and enhances resolution of complex genetic architecture. Specifically, we integrated publicly available GWAS summary statistics from multiple ASCVD-related diseases (Malik et al. 2018, Sakaue et al. 2021, Aragam et al. 2022, Roychowdhury et al. 2023), enabling the estimation of single-nucleotide polymorphism (SNP)-level associations with the latent ASCVD construct.

Over recent years, transcriptome-wide association studies (TWAS) have emerged as a powerful analytical framework for identifying gene-level associations with complex traits by leveraging genetically imputed gene expression (Gamazon et al. 2015). This approach is grounded in the observation that a substantial proportion of GWAS-identified risk variants colocalize with expression quantitative trait loci (eQTLs), thereby implicating regulatory mechanisms that influence gene expression (Hormozdiari et al. 2016). By modelling the genetic component of gene expression, TWAS not only enhances statistical power but also reduces susceptibility to environmental confounding and mitigates reverse causation, thereby improving the fine-mapping of functionally relevant genomic regions (Gusev et al. 2016, Mancuso et al. 2017). Conventional TWAS analyses have typically been performed on a tissue-by-tissue basis, an approach that may overlook shared regulatory architecture across tissues (Liu et al. 2017). However, accumulating evidence indicates that a considerable number of eQTLs exert multi-tissue regulatory effects (Battle et al. 2017). To address this limitation, recent methodological advances have led to the development of the Unified Test for Molecular Signatures (UTMOST)-a cross-tissue TWAS framework that enables simultaneous gene-level association testing across multiple tissues (Hu et al. 2019). By applying a group-lasso penalty to SNP effect sizes across tissues, UTMOST effectively integrates eQTLs with multi-tissue regulatory influence while preserving tissue-specific signals of biological significance. This cross-tissue strategy has proven instrumental in pinpointing susceptibility genes for a range of complex disorders, including diabetes (Fu et al. 2026), obesity and metabolic disorders (Fu et al. 2025), underscoring its utility in elucidating the genetic architecture of human disease.

To identify genes associated with atherosclerotic cardiovascular disease (ASCVD), we performed a cross-tissue transcriptome-wide association study (TWAS) using the UTMOST framework, integrating genomic SEM-derived latent phenotype data with expression quantitative trait locus (eQTL) information from the Genotype-Tissue Expression (GTEx) project version 8. To further refine tissue-specific associations, we applied the Functional Summary-based Imputation (FUSION) method (Liufu et al. 2024), a computational tool that imputes genetically regulated gene expression and tests its association with complex traits using GWAS summary statistics. In parallel, we conducted gene-level association analyses using MAGMA (Leeuw et al. 2015), which aggregates SNP-level signals within genes and evaluates their joint association with the ASCVD phenotype. Candidate genes were defined as those showing consistent evidence of association across all three analytical approaches-UTMOST, FUSION, and MAGMA. To generate hypotheses about the biological relevance and translational potential of the identified genes, we performed a comprehensive series of downstream analyses that should be interpreted as hypothesis-generating rather than definitive mechanistic validation. These included Mendelian randomizations (MR) and colocalization to prioritize candidate causal relationships; phenome-wide association studies (PheWAS) to explore pleiotropic effects; GeneMANIA-based network analysis and pathway enrichment to characterize functional interactions; and evaluations of tissue- and cell-type specificity. Additionally, we integrated gene structure predictions with chemical-disease interaction profiling to further explore mechanistic implications. Together, these multiple convergent statistical approaches provided multi-layered hypotheses into the biological characteristics and potential of the candidate genes implicated in ASCVD.

Methods

Figure 1 provides a schematic overview of the analytical workflow implemented in this study. All analyses were conducted using publicly available summary-level statistics and gene expression datasets, the details and access information for which are provided in the ‘Data Availability’ section. Ethical approval for the original studies contributing to these datasets was obtained by the respective institutional review boards, and all procedures were performed in accordance with relevant ethical guidelines and regulations.

Fig. 1.

Flowchart of this study design. Abbreviations: ASCVD, Atherosclerotic Cardiovascular Disease; SEM, Structural equation modelling; CHD, Coronary Heart Disease; TIA, Transient Ischemic Attack; PAD, Peripheral Arterial Disease; AAA, Abdominal Aortic Aneurysm; GTEx, Genotype-Tissues Expression Project; TWAS, transcriptome-wide association studies; UTMOST, unified test for molecular signatures; FUSION, functional summary-based imputation; MAGMA, multi-marker Analysis of GenoMic Annotation; MR, Mendelian randomization; TSEA, GTEx Tissue Specific Expression Analysis; CSEA, Cell-type-specific enrichment analysis

Data sources for ASCVD and genomic SEM

To investigate the genetic architecture of atherosclerotic cardiovascular disease (ASCVD), we applied genomic structural equation modeling (Genomic SEM) using the GenomicSEM R package (v.0.0.5) (Grotzinger et al. 2019). This multivariate framework integrates genome-wide association study (GWAS) summary statistics for five ASCVD-related traits: coronary heart disease (CHD) (Aragam et al. 2022), stroke (Malik et al. 2018), transient ischemic attack (TIA) (Kurki et al. 2023), abdominal aortic aneurysm (AAA) (Roychowdhury et al. 2023), and peripheral artery disease (PAD) (Sakaue et al. 2021). A total of 3,914,019 single-nucleotide polymorphisms (SNPs) were included in the analysis to identify genetic variants associated with latent ASCVD susceptibility. The summed sample size across the five GWAS studies is approximately 3.8 million individuals; this aggregate number is reported for descriptive purposes and does not represent the effective sample size of the genomic SEM analysis, which depends on model specification and LD score regression. Detailed information on the univariate GWAS datasets used in Genomic SEM is provided in Table S1 and Table S2, and model-fitting parameters are summarized in Table S3 and Table S4. For multi-ancestry GWAS (e.g., CAD and stroke), we restricted analysis to individuals of European ancestry only to ensure population homogeneity. All input summary statistics were therefore derived from European-ancestry populations.

Model specification and dimensionality assessment. Before performing confirmatory factor analysis, we evaluated the suitability of the genetic covariance matrix for factor analysis using the Kaiser‑Meyer‑Olkin (KMO) measure. The overall KMO was 0.69, indicating moderate common variance among the five traits, which does not strongly favour a complex multi‑factor structure. Based on the genetic covariance matrix among the five traits and applying the maximum number of factors formula, we determined that the maximum number of plausible factors is two. Guided by the factor loadings from exploratory factor analysis, we then assigned specific traits to each factor: CHD and PAD loaded on Factor 1, while stroke and TIA loaded on Factor 2. Nevertheless, given our primary goal of discovering shared genetic signals across all five clinically relevant ASCVD phenotypes, we proceeded with a one‑factor common factor model including all five traits. Model fit was evaluated using CFI, SRMR, χ², and AIC. As a sensitivity analysis (Table S21), we also tested a two‑factor model (CHD + PAD vs. stroke + TIA). Because the factor correlation was high (r ≈ 0.80) and the one‑factor model already met good fit criteria, we retained the one‑factor model as the primary, more parsimonious and comprehensive representation.

Quality control of univariate input GWAS

Prior to genomic SEM, we performed standardized quality control on each input GWAS summary statistic. Samples with genotype missing rate exceeding 5% were excluded. The major histocompatibility complex (MHC) region (chr6:25–35 Mb) was removed due to its complex linkage disequilibrium structure. For SNP-level filtering, we retained only autosomal variants with minor allele frequency (MAF) ≥ 0.01, effect size not equal to zero, and consistent allele alignment with the 1000 Genomes Phase 3 European reference panel. SNPs with allele mismatches or ambiguous strand orientation were discarded. For LD score regression and covariance matrix estimation, we further restricted to HapMap3 variants to minimize estimation bias. These procedures ensured reliable multivariate GWAS input for genomic SEM. Genomic SEM and LD score regression inherently account for potential sample overlap among input GWAS by estimating sampling covariances, thus mitigating bias due to shared individuals.

eQTL data acquisition

To obtain tissue-specific gene expression data, we utilized the Genotype-Tissue Expression (GTEx) Project, Version 8 dataset (Lonsdale et al. 2013), which comprises gene expression profiles across 49 distinct tissue types derived from 838 post-mortem donors (https://ftp.ebi.ac.uk/pub/databases/spot/eQTL/imported/GTEx_V8/ge/). Sample sizes varied substantially across tissues, ranging from 73 in the renal cortex to 706 in skeletal muscle, reflecting the inherent heterogeneity of tissue availability in post-mortem collections.

Cross-tissue analyses for ASCVD

To systematically investigate gene-level associations with ASCVD at the organismal level, we applied the cross-tissue UTMOST method (Hu et al. 2019)(https://github.com/Joker-Jerome/UTMOST). This framework integrates expression data across multiple tissues, thereby enhancing statistical power by leveraging tissues with increased trait heritability and improved imputation accuracy (Hu et al. 2019, Lo Faro et al. 2024). Gene-trait association statistics were subsequently aggregated using the generalized Berk-Jones (GBJ) test, which accounts for the covariance structure derived from single-tissue summary statistics (Hu et al. 2019, Sun et al. 2019). Associations with a false discovery rate (FDR) below 0.05 were considered statistically significant after multiple-testing correction.

Single tissue analyses for ASCVD

To systematically identify genes associated with ASCVD, we performed TWAS using the FUSION tool (Gusev et al. 2016)(http://gusevlab.org/projects/fusion/). This framework integrates GWAS summary statistics for ASCVD with expression quantitative trait locus (eQTL) data from 49 tissues in the GTEx version 8 reference panel. For each gene, linkage disequilibrium (LD) between SNPs in the predictive model and the broader GWAS locus was estimated using the 1000 Genomes Project European reference samples. FUSION evaluates a spectrum of prediction models-including LASSO, Elastic Net, BSLMM, BLUP, and a top SNP-based model-to identify the best-performing model for imputing genetic components of gene expression based on cross-validated performance. The model with the highest predictive accuracy was selected to derive gene expression weights, which were then integrated with ASCVD GWAS Z-scores to test for gene-trait associations. Genes achieving a FDR < 0.05 in both the cross-tissue UTMOST and single-tissue FUSION analyses were prioritized as candidate susceptibility genes for ASCVD.

Conditional and joint analysis

To delineate conditionally independent genetic signals within associated loci, we applied the conditional and joint analysis (COJO) module implemented in the FUSION framework (Gusev et al. 2016). This post-GWAS step is essential for refining association signals by accounting for LD among markers, thereby clarifying the genetic architecture underlying trait variability (Liao et al. 2019). In this study, the genes identified by both UTMOST and FUSION (61 significant genes for ASCVD) were selected for COJO analysis. COJO analysis was performed specifically in the tissues where FUSION had initially identified significant gene-level associations. Genes that remained significant after conditioning on the lead variant were classified as jointly associated, whereas those that lost significance were considered marginally associated. Multiple testing correction was applied using the false discovery rate (FDR) at this stage. A total of 61 genes significant in both UTMOST and FUSION analyses were selected as the input set for this conditional analysis.

Workflow transparency

To ensure workflow transparency, we quantitatively tracked the number of SNPs and genes retained at each major analytical step. After removing MHC region, low-frequency variants (MAF < 0.01), and non-HapMap3 variants for five GWAS, 3,914,019 SNPs were included in the analysis to identify genetic variants associated with latent ASCVD susceptibility and finally used for LD score regression and covariance estimation. In cross-tissue UTMOST, 20,609 genes were tested across 49 tissues, of which 271 remained significant at FDR < 0.05 (Table S5). In single-tissue FUSION, approximately 19,100 genes were tested per tissue (varying slightly due to eQTL availability), yielding 2,637 unique gene-tissue associations (FDR < 0.05, Table S6). MAGMA tested 18,006 genes and identified 2,059 significant associations (FDR < 0.05, Table S9). The intersection of UTMOST and FUSION produced 61 genes (Table S7 and Table S8); further intersection with MAGMA yielded 14 high-confidence genes (Table S10). After conditional analysis, four genes (ARVCF, GFPT1, NFU1, USP39) were prioritized for downstream investigation. Regarding code availability, all project-specific scripts and exact analysis parameters are provided in the Code Availability section; unless otherwise specified in the Methods, default parameters of the cited software packages were used.

Gene analysis by MAGMA

To complement the transcriptome-wide association analyses, we conducted gene-level association testing using MAGMA (v1.10). This approach aggregates SNP-level association statistics from GWAS summary data into a composite gene score, thereby quantifying the overall evidence of association between each gene and the phenotype of interest (Leeuw et al. 2015, Leeuw et al. 2018). SNP annotation and LD structure were derived from Phase 3 of the 1000 Genomes Project (European ancestry reference panel). Detailed descriptions of the parameters and analytical procedures are available in the original MAGMA documentation (Leeuw et al. 2015).

Mendelian randomization and colocalization analysis

To investigate potential causal relationships between gene expression and ASCVD, we performed two-sample MR analyses using the “TwoSampleMR” package in R (Hemani et al. 2018). Cis-expression quantitative trait loci (cis-eQTLs) were employed as instrumental variables (IVs), with gene expression levels serving as the exposure and ASCVD GWAS summary statistics as the outcome. Instrumental variables were initially selected based on genome-wide significance (P < 5 × 10^− 8) and subsequently pruned for linkage disequilibrium (LD clumping, r² < 0.001) to retain only independent SNPs (Yuan et al. 2023). To ensure the validity of the instrumental variable assumptions, we excluded cis-eQTLs with potential pleiotropic effects, as many eQTLs may influence multiple tissues or traits, thereby violating the exclusion restriction criterion of MR. Instrument strength was assessed by calculating the F‑statistic for each cis‑eQTL; all instrumental variables had F‑statistics > 10, confirming the absence of weak instrument bias. The Wald ratio method was applied as the primary MR estimator, providing unbiased causal estimates under the assumption of valid instruments. To further guard against pleiotropy, we required Bayesian colocalization evidence (PP.H4 > 0.6) as a complementary criterion for prioritizing candidate causal relationships, thereby reducing reliance on MR alone. The MR effect estimation was facilitated using the Wald ratio method, with statistical significance established at p < 0.0027 for ARVCF (18 tissues), p < 0.0083 for CGREF1, DPYSL5 and GNB1L (6 tissues), p < 0.01 for COL4A4 (5 tissues), p < 0.016 for CXCR2 and HTT (3 tissues), p < 0.05 for DUSP11 (1 tissue), p < 0.005 for GFPT1 (10 tissue), p < 0.0015 for GNLY (33 tissue), p < 0.0035 for NDNF and USP39 (14 tissue), p < 0.0041 for NFU1 (12 tissue), and p < 0.025 for WDR43 (16 tissues) considering multiple testing.

To assess whether the same causal variants underlie both the ASCVD GWAS signals and the expression regulation of candidate genes, we performed Bayesian colocalization analysis using the “coloc” package in R (Giambartolomei et al. 2014). This method evaluates the posterior probability (PP) for five mutually exclusive hypotheses regarding the sharing of causal variants between two traits, with PP.H4 representing the scenario in which both traits share a single causal variant (Giambartolomei et al. 2014). A PP.H4 value greater than 0.6 was considered as suggestive evidence of a shared causal variant, whereas PP.H4 > 0.8 or > 0.9 would indicate stronger support. We also report PP.H3 (probability of distinct causal variants) for completeness. All colocalization results should be interpreted as hypothesis-generating rather than confirmatory.

PheWAS analysis of ASCVD candidate genes

To systematically evaluate the pleiotropic profile and potential safety implications of the prioritized genes, we performed a phenome-wide association study (PheWAS) using the AstraZeneca PheWAS Portal (https://azphewas.com/) (Wang et al. 2021). This public resource provides comprehensive gene-phenotype association results derived from exome sequencing and linked phenotypic data of approximately 500,000 UK Biobank participants (Wang et al. 2021). We specifically queried associations for the four novel ASCVD candidate genes, ARVCF, GFPT1, NFU1 and USP39, cross a broad spectrum of binary and quantitative traits. This analysis aimed to uncover potential horizontal pleiotropy and identify any adverse phenotype associations that may inform the translational potential and therapeutic safety of targeting these genes.

GeneMANIA analysis

To explore the functional context and potential biological roles of the candidate genes, we employed the GeneMANIA platform (https://genemania.org/) (Mostafavi et al. 2008). This platform integrates a wide range of publicly available datasets, including protein-protein genetic interactions, co-expression data, and pathway annotations, to generate hypotheses about gene function ⁸¹. By constructing interaction networks for our prioritized genes, GeneMANIA enabled the identification of enriched biological processes and pathway crosstalk, thereby providing insight into their potential roles within the cellular and molecular frameworks relevant to ASCVD.

Pathway enrichment analysis

To functionally annotate the prioritized genes and identify overrepresented biological pathways, we performed enrichment analyses using the g: Profiler tool (Raudvere et al. 2019). Gene sets were queried against the Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) databases. To correct for multiple testing, we applied the g: SCS (Set Counts and Sizes) method, which is implemented in g: Profiler to account for the inherent dependencies among functionally related gene sets and provides a balanced correction for term overlap. Enriched terms with an adjusted P-value below the significance threshold were considered for further interpretation.

GTEx tissue specific expression analysis (TSEA)

To systematically evaluate tissue specificity of the ASCVD-associated genes, we performed tissue-specific expression analysis (TSEA) using RNA-seq data from the Genotype-Tissue Expression (GTEx) project, which provides comprehensive gene expression profiles across multiple tissues derived from post-mortem donors (Dougherty et al. 2010). This analysis aimed to determine whether the prioritized genes exhibited significant enrichment in particular tissues compared to the genomic background. Hypergeometric tests were applied to assess overrepresentation, and multiple testing correction was performed using the Benjamini-Hochberg method to control the FDR below 0.05.

Cell-type-specific enrichment analysis (CSEA)

To systematically assess enrichment of ASCVD-associated genes within specific cell types, we performed cell-type-specific enrichment analysis (CSEA) using the WebCSEA platform (Dai et al. 2022). This resource integrates expression signatures derived from 11 single-cell RNA-seq datasets, encompassing over 5.5 million cells across 111 tissues and 1,355 tissue-cell-type (TC) categories, as compiled by Dai et al. 2022. In their workflow, low-expression genes were filtered out, and the “deTS” t-statistic-based method was applied to identify signature genes for each cell type, defined as those ranking in the top 5% of t-statistic scores. To evaluate whether our prioritized ASCVD genes were overrepresented in specific cell types, we applied Fisher’s exact test comparing the candidate gene list against each cell-type signature set.

Structure predictions and gene-chemical-disease analysis for candidate genes for ASCVD

To gain structural and functional insights into the four prioritized novel genes (ARVCF, GFPT1, NFU1 and USP39), we first consulted the GeneCards database (https://www.genecards.org/) to retrieve their comprehensive gene structures and genomic annotations. Subsequently, to explore their potential interactions with environmental chemicals and to assess their broader mechanistic landscape, we interrogated the Comparative Toxicogenomics Database (CTD) database (https://ctdbase.org/) (Davis et al. 2025). CTD is a public resource that manually curates and integrates data on chemical-gene/protein interactions, chemical-disease relationships, and gene-disease associations, thereby enabling predictions about how environmental exposures may influence disease pathogenesis via specific genes. By querying CTD for each candidate gene, we aimed to identify known or predicted chemical interactants that could modulate their activity and to infer potential links to ASCVD-relevant pathways.

Results

Utilizing Genomic SEM to ASCVD

To investigate the genetic overlap among the five ASCVD-related traits, we first performed linkage disequilibrium score regression (LDSC). This analysis revealed positive genetic correlations between CHD, stroke, TIA, PAD, and AAA (Tables S1 and S2). To capture this shared genetic architecture, we applied genomic SEM. A common-factor model demonstrated excellent fit to the empirical genetic covariance matrix derived from the five input GWAS (Comparative Fit Index [CFI] = 0.97; Standardized Root Mean Square Residual [SRMR] = 0.067; Tables S3 and S4), providing robust statistical evidence for a latent “multivariate ASCVD” genetic factor. We subsequently extended this SEM framework to incorporate individual variant effects, conducting a multivariate GWAS that quantified SNP-level associations with the shared ASCVD factor across 3,914,019 genetic variants.

To further evaluate the factor structure, we also tested a two-factor confirmatory factor model in which CHD and PAD loaded on one factor and stroke and TIA on another, with AAA allowed to cross-load based on prior biological knowledge. This model yielded comparable fit statistics (CFI = 0.98, SRMR = 0.051; Table S21). Given the principle of parsimony and the excellent fit of the common-factor model, we retained the single-factor solution for subsequent multivariate GWAS and downstream analyses.

Identification of ASCVD candidate genes across multi-tissue and single-tissue TWAS analyses

The cross-tissue TWAS (UTMOST) identified 625 genes with nominal significance (P < 0.05; Table S5). After FDR correction (FDR < 0.05), 271 genes remained significantly associated with ASCVD (Fig. 2A and B). Single-tissue TWAS (FUSION) identified 2,637 significant gene-tissue associations (FDR < 0.05; Table S6). Intersecting both approaches yielded 61 high-confidence genes (Fig. 2A and B; Table S7).

Fig. 2.

The distribution of genes identified by different methods. A Upset plot illustrating the overlap of genes identified by various methods as being associated with ASCVD; The y-axis represents the number of genes. B Venn diagram of the overlap of genes identified by various methods. C Heatmap of fusion TWAS.P values for 14 identified ASCVD genes (intersected by UTMOST, FUSION and MAGMA) in different tissues. Abbreviations: ASCVD, Atherosclerotic Cardiovascular Disease; TWAS, transcriptome-wide association studies; UTMOST, unified test for molecular signatures; FUSION, functional summary-based imputation; MAGMA, multi-marker Analysis of GenoMic Annotation

Conditional and joint multiple-SNP analysis identifies independent and interacting genetic signals for ASCVD

COJO analyses on the 61 genes identified independent signals for ARVCF in artery aorta (Fig. 3A), GFPT1 in heart atrial appendage (Fig. 3B), and USP39 in heart left ventricle (Fig. 3D). Notably, in adipose subcutaneous tissue, conditioning on NFU1 expression increased the GFPT1 signal, reflecting complex local LD rather than direct biological interaction (Fig. 3C). Full results in Table S8.

Fig. 3.

The fusion-cojo analysis plots for identified genes for ASCVD. A Regional plot of ARVCF gene in artery aorta for ASCVD. B Regional plot of GFPT1 gene in heart atrial appendage for ASCVD. C Regional plot of NFU1 gene in adipose subcutaneous for ASCVD. D Regional plot of USP39 gene in heart left ventricle for ASCVD. The top panel highlights all genes in the region. COJO analysis is used for tests of individual coefficients. Each test is two-sided and the P value is evaluated with false discovery rate. Abbreviations: ASCVD, Atherosclerotic Cardiovascular Disease. The marginally associated TWAS genes are shown in blue, and the jointly significant genes are shown in green. The bottom panel shows a regional Manhattan plot of GWAS data before (grey) and after (blue) conditioning on the predicted expression of the green genes

Here we highlight the results for four genes (ARVCF, GFPT1, NFU1 and USP39) that were consistently identified across all three analytical approaches. Of note, ARVCF has prior evidence in CAD GWAS (Aragam et al. 2022) and functional studies (Barbera et al. 2025), whereas USP39 is newly prioritized in the context of the composite ASCVD phenotype. In artery aorta, the TWAS signal for ARVCF remained significant after conditioning on the expression of neighboring genes, indicating an association independent of local LD (Fig. 3A). Similarly, GFPT1 in heart atrial appendage and USP39 in heart left ventricle retained their significant associations after conditioning, supporting their independence from LD-confounded effects (Fig. 3B and D). Notably, in adipose subcutaneous tissue, the TWAS signal for GFPT1 became more significant when conditioned on NFU1 expression, reflecting complex local LD or correlated predicted expression rather than a direct biological interaction (Fig. 3C). Full results of COJO analyses across all 61 genes and their respective tissues are provided in Table S8.

Gene association and cross-tissue integration identify candidate genes for ASCVD

MAGMA identified 2,059 significant gene-level associations (FDR < 0.05; Fig. 2A and B; Table S9). Integrating MAGMA with the 61 UTMOST-FUSION genes yielded 14 high-confidence genes (Fig. 2A and B), including ARVCF, GFPT1, NFU1, and USP39 (Fig. 2C; Table S10). ARVCF has been previously implicated in CAD GWAS (Aragam et al. 2022) and experimentally validated in human aortic smooth muscle cells (Barbera et al. 2025); USP39 has not been previously reported in association with the composite ASCVD phenotype. However, such broad expression may reflect general cellular functions rather than ASCVD-specific biology, and we therefore rely on tissue enrichment and colocalization evidence to infer disease-relevant specificity.

Causal and colocalization analyses uncover tissue-specific associations for candidate genes in ASCVD

To investigate the causal roles of the 14 candidate genes (intersection by UTMOST, FUSION and MAGMA) in ASCVD, we employed MR and colocalization analyses (Table S11 and Fig. 4). We focus on four genes (ARVCF, GFPT1, NFU1 and USP39), which showed detectable expression across multiple tissues; tissue-specific enrichment and colocalization evidence were subsequently used to prioritize disease-relevant contexts. The ARVCF gene is located on chromosome 22q11.21, and the results of FUSION demonstrate its significant expression in 20 tissues for ASCVD (Table S10). The MR analysis identified significant associations consistent with potential causal effects in 11 tissues (p value corrected by multiple testing < 0.05) (Fig. 4A and Table S11), with an OR (95%CI) of 0.97 (0.96, 0.98) in artery aorta, 0.95 (0.93, 0.96) in artery coronary and 1.05 (1.03, 1.07) in artery tibial for ASCVD. Colocalization analysis provided supportive evidence, with PP.H4 values of 0.99, 0.98 and 0.99 for artery aorta, artery coronary and artery tibial, respectively (see Table S12 for PP.H3 values and full colocalization probabilities). Notably, rs9606203 emerged as the most significant colocalization locus for ASCVD in artery aorta (Fig. 4B).

Fig. 4.

The results of Mendelian Randomization and colocalization analysis between candidate genes and ASCVD. A The MR results confirmed the causal associations between three candidate genes and ASCVD. Wald ratio test is used for tests of individual coefficients. Each test is two-sided and the original P value is reported with no multiple comparisons. Horizontal lines indicate the 95% confidence intervals around the effect sizes. The GTEx’s final dataset (V8) contains DNA data from 838 postmortem donors and 17,382 RNA-seq across tissue sites and cell lines. B The results of colocalization analysis between ARVCF and ASCVD in artery aorta. C The results of colocalization analysis between GFPT1 and ASCVD in heart atrial appendage. D The results of colocalization analysis between NFU1 and ASCVD in adipose subcutaneous. E The results of colocalization analysis between USP39 and ASCVD in artery aorta. Colocalization analysis is used for tests of posterior probability (PP). A PP.H4 greater than 0.6 indicated suggestive evidence of shared causal variants. Abbreviations: ASCVD, Atherosclerotic Cardiovascular Disease; OR, odds ratio

GFPT1, located on chromosome 2p13.3, was notably correlated with ASCVD in 19 tissues according to FUSION analysis (Table S10). MR analysis provided evidence consistent with a causal effect between GFPT1 and ASCVD in 8 tissues (p value corrected by multiple testing < 0.05) (Fig. 4A and Table S11). Notably, the esophagus muscularis showed a significant MR estimate (OR = 0.94, p-value = 8.45E-07), but colocalization support was only suggestive (PP.H4 = 0.64), and the biological relevance of this tissue to ASCVD is uncertain. In contrast, heart atrial appendage exhibited stronger colocalization evidence (PP.H4 = 0.87; Table S12), making it a more credible candidate tissue for GFPT1 function in ASCVD. Rs7424361 emerged as the most significant colocalization locus for ASCVD in this tissue (Fig. 4C). Additionally, FUSION analysis revealed that the NFU1 gene is a significant contributor to ASCVD in 15 tissues (Table S10). MR analyses provided evidence consistent with a causal effect in 5 tissues (p < 0.05 after correction by multiple testing) (Fig. 4A and Table S11), with the significant association found in the adipose subcutaneous (OR = 1.11, p value = 8.17E-08). Colocalization analysis further supported this with PP.H4 values of 0.95 for adipose subcutaneous (Table S12). rs7422275 was the most significant colocalization locus for ASCVD in adipose subcutaneous (Fig. 4D).

Similarly, MR analysis identified significant causal associations of USP39 with ASCVD in 4 tissues (p < 0.05 after correction by multiple testing) (Fig. 4A and Table S11). Notably, rs2232748 emerged as the most significant colocalization locus for ASCVD in artery aorta (Fig. 4E).

PheWAS analysis reveals gene-phenotype associations for ASCVD

For ARVCF, GFPT1, and NFU1, we observed significant associations with a range of binary phenotypes, particularly those related to cardiovascular and musculoskeletal diseases (Fig. 5A and C, and 5E). These genes also showed notable associations with quantitative traits, including proteomic biomarkers, musculoskeletal measurements, laboratory findings, and health service utilization indicators (Fig. 5B, D and F, and 5H). These findings suggest potential horizontal pleiotropy and raise considerations regarding possible on-target side effects that may inform future therapeutic development. In contrast, USP39 variants exhibited no significant associations with the binary phenotypes examined (Fig. 5G), indicating a potentially more specific role in ASCVD pathogenesis.

Fig. 5.

The Manhattan plot of phenome-wide association study and GeneMania gene network for candidate genes. A The Manhattan plot of phenome-wide association between ARVCF and UKB binary traits. B The Manhattan plot of phenome-wide association between ARVCF and UKB continuous traits. C The Manhattan plot of phenome-wide association between GFPT1 and UKB binary traits. D The Manhattan plot of phenome-wide association between GFPT1 and UKB continuous traits. E The Manhattan plot of phenome-wide association between NFU1 and UKB binary traits. F The Manhattan plot of phenome-wide association between NFU1 and UKB continuous traits. G The Manhattan plot of phenome-wide association between USP39 and UKB binary traits. H The Manhattan plot of phenome-wide association between USP39 and UKB continuous traits. Generalized linear model is used for phenome-wide association study. Each test is two-sided and the original P value is reported with no multiple comparisons. I GeneMania gene network of ARVCF as the core. J GeneMania gene network of GFPT1 as the core. K GeneMania gene network of NFU1 as the core. L GeneMania gene network of USP39 as the core. Abbreviations: UKB, UK Biobank

GeneMANIA network analysis reveals pathway enrichments for ASCVD-associated genes

The resulting network positioned ARVCF centrally, with its associated genes primarily enriched in pathways related to cell-cell junction organization, synapse organization, and cell-cell adhesion (Fig. 5I; Table S13). A second network centered on GFPT1 (Fig. 5J) was predominantly enriched in metabolic processes, including UDP-N-acetylglucosamine metabolism, amino sugar biosynthesis, nucleotide-sugar metabolism, regulation of cellular carbohydrate metabolism, and ER-nucleus signaling (Table S14). The NFU1-centered network (Fig. 5K) showed significant enrichment in pathways related to metallo-sulfur cluster assembly, consistent with its known role in iron-sulfur cluster biogenesis (Table S15). Finally, the network centered on USP39 (Fig. 5L) was enriched for spliceosomal complex components, including the tri-snRNP complex and small nuclear ribonucleoprotein particles, implicating it in RNA splicing regulation (Table S16).

Together, these network analyses reveal distinct biological roles for the four genes-ranging from cell adhesion (ARVCF) and metabolic regulation (GFPT1) to mitochondrial function (NFU1) and RNA processing (USP39)-highlighting the functional diversity of genetic factors contributing to ASCVD susceptibility. These findings are exploratory and should be interpreted as hypothesis-generating.

Pathway enrichment analysis uncovers biological functions of ASCVD-associated genes

The 14 ASCVD-associated genes showed significant enrichment in GO terms related to protein binding, glutamine metabolic process, cell growth, and nucleoplasm (Fig. 6A). KEGG pathway analysis further highlighted enrichment in nucleotide sugar biosynthesis and amino sugar and nucleotide sugar metabolism (Fig. 6A). These findings suggest that perturbations in metabolic pathways-particularly those involving amino sugar and nucleotide sugar metabolism-may play a contributory role in ASCVD pathogenesis, complementing the well-established roles of lipid metabolism and inflammation, but this interpretation remains speculative and requires experimental validation.

Fig. 6.

Biological function of the identified ASCVD genes. A Results of the functional enrichment analysis for identified 14 ASCVD-related genes (intersected by UTMOST, FUSION and MAGMA). B Tissue-specific gene enrichment results of identified 14 ASCVD genes. Bars with red color means significantly enriched DEG sets. C Cell-type-specific gene enrichment plot of identified 14 ASCVD related genes. The structural prediction by AlphaFold for ARVCF gene (D), GFPT1 gene (E), NFU1 gene (F) and USP39 gene (G). Chemical interactions for ARVCF gene (H), GFPT1 gene (I), NFU1 gene (J) and USP39 gene (K). The x-axis represents the number of interactions from the CTD (Comparative Toxicogenomics Database). The y-axis indicates the chemicals. The assessment of functional enrichment for the genes was assessed using the hypergeometric test. The g: SCS method was used for multiple testing correction with two-sided. Abbreviations: ASCVD, Atherosclerotic Cardiovascular Disease; BP, Biological process; GO, Gene Ontology; MF, Molecular function; CC, Cellular Component; KEGG, Kyoto Encyclopedia of Genes and Genomes; DEG, Differentially Expressed Genes; UTMOST, unified test for molecular signatures; FUSION, functional summary-based imputation; MAGMA, multi-marker Analysis of GenoMic Annotation

Tissue and cell-type specificity reveal key players in ASCVD

We performed an integrative analysis combining bulk RNA-seq data from the GTEx project (Lonsdale et al. 2013) with single-cell transcriptomic profiles from 11 published datasets (Dai et al. 2022). This approach revealed significant enrichment of the 14 ASCVD-associated genes in specific tissues, most notably the heart atrial appendage and the cervical c-1 segment of the spinal cord (Fig. 6B). At the cellular level, we observed significant enrichment in multiple immune and progenitor cell populations, including erythroid progenitor cells, alpha-beta T cells, classical monocytes, plasma cells, and skeletal muscle satellite stem cells (Fig. 6C). Collectively, these findings implicate these genes in immune regulation, tissue repair, and local homeostatic maintenance-processes increasingly recognized as important contributors to vascular health and disease.

Structure predictions, chemical interactions and disease pathways of identified genes in ASCVD

To explore potential structural and chemical features of the prioritized genes (ARVCF, GFPT1, NFU1 and USP39) in a hypothesis-generating manner, we next investigated their protein structures, potential chemical interactions, and associated disease pathways.

Protein structures predicted by AlphaFold3 for all four genes demonstrated high confidence, with the majority of residues achieving a predicted local distance difference test (pLDDT) score above 90, corresponding to a deep blue color in the structural model (Fig. 6D and G). Only a few peripheral regions, located away from missense mutation sites, showed pLDDT scores below 70. These predicted models provide a hypothesis-generating structural foundation for future functional studies.

Using the Comparative Toxicogenomics Database (CTD) (Davis et al. 2025), we explored potential chemical-gene interactions for the four genes. This analysis revealed a diverse set of potential chemical interactants. For example, ARVCF showed interactions with benzopyrene, acetaminophen, ozone, and acrolein (Fig. 6H). GFPT1 interacted with bisphenol A, cyclosporine, glucose, and tetrachlorodibenzodioxin (Fig. 6I). NFU1 demonstrated potential interactions with bisphenol A and tetrachlorodibenzodioxin (Fig. 6J), while USP39 showed interactions with bisphenol A, cisplatin, tetrachlorodibenzodioxin, ozone, and sodium arsenite (Fig. 6K). Further analysis of the underlying gene-chemical-disease networks linked these genes to a broad spectrum of pathophysiological conditions, including weight dynamics, liver and kidney diseases, inflammation, fatty liver, and various cardiovascular and metabolic disorders (Tables S17-S20). Notably, GFPT1 was highlighted as a potential biomarker for obesity, muscular diseases, and myasthenic syndrome (Table S18), whereas USP39 was associated with colonic neoplasms (Table S20). In an exploratory context, these findings suggest that beyond their role in ASCVD susceptibility, these genes may participate in broader biological processes, but these hypotheses require direct experimental testing. No specific cardiovascular drug or variant interaction is implied by these analyses.

Discussion

This study provides a comprehensive, multi-layered investigation into the genetic architecture and molecular pathophysiology of ASCVD and its major clinical manifestations, including CHD, stroke, TIA, PAD, and AAA. By integrating GWAS, genomic SEM, cross-tissue and single-tissue TWAS, and MAGMA-based gene-level analyses, we identified 14 genes robustly associated with ASCVD susceptibility. These genes were further prioritized using MR and colocalization analyses as candidate susceptibility genes. These statistical approaches provide hypothesis-generating evidence rather than definitive causal proof. Collectively, our findings suggest potential roles for the identified genes in ASCVD, but all interpretations should be considered exploratory and require experimental validation. This integrative framework provides a statistical prioritization resource for future functional studies.

Recent genetic studies have increasingly focused on elucidating the genetic relationships among various cardiovascular disease phenotypes. Using genomic SEM, we observed significant genetic covariances among these phenotypes, suggesting the presence of shared genetic factors. This observation aligns with the hypothesis proposed by Rheen et al. (2019) that genetic correlations among certain disease phenotypes may reflect common genetic determinants, which can exert broad effects on disease etiology, pathogenesis, and clinical outcomes (Rheenen et al. 2019). In our analysis, this pattern was particularly pronounced for CHD, stroke, and PAD, indicating that these conditions are not genetically isolated but rather interconnected through shared genetic influences. By integrating GWAS data across diverse populations and environmental contexts, we identified genetic effects that transcend traditional diagnostic boundaries, further reinforcing the central role of genetic factors in the mechanistic architecture of multiple cardiovascular diseases. These findings are consistent with and extend recent work by Dichgans et al. (2014) (Dichgans et al. 2014) and Haritala et al. (2021) (Hartiala et al. 2021), who demonstrated significant overlap of common risk loci between CHD and stroke. Collectively, our results underscore the value of multivariate genetic approaches in dissecting the complex architecture of cardiovascular disease and highlight the biological relevance of shared genetic pathways across traditionally distinct clinical entities.

A central challenge in the post-GWAS era is that the majority of disease-associated variants reside in non-coding regions of the genome (Maurano et al. 2012), with only a small fraction having been functionally annotated (Gamazon et al. 2015). This observation underscores the critical role of gene regulation in disease susceptibility, a notion further supported by the significant enrichment of eQTLs among GWAS-identified variants (Nicolae et al. 2010). TWAS address this gap by integrating eQTL data with GWAS summary statistics to test for association between genetically regulated gene expression and complex traits (Gamazon et al. 2015). By applying a multi-tissue TWAS framework to ASCVD, our study not only provides evidence for shared genetic architecture and potential causal relationships but also indicates the utility of this approach in prioritizing putatively functional genes and uncovering tissue-specific regulatory mechanisms underlying disease pathogenesis.

This study advances previous work by applying the UTMOST framework for cross-tissue transcriptome-wide association studies (TWAS), which represents a significant methodological improvement over conventional single-tissue approaches. By integrating gene expression data across multiple tissues, UTMOST enables a more comprehensive characterization of gene–trait associations, thereby enhancing statistical power and facilitating the detection of regulatory effects that may be missed in tissue-restricted analyses (Hu et al. 2019). Through this robust integrative framework, followed by complementary prioritization approaches, we identified 14 genes robustly associated with ASCVD susceptibility. Subsequent bioinformatics analyses revealed that these candidate genes are primarily involved in biological processes such as amino sugar biosynthesis, nucleotide-sugar metabolism, and cell growth. Moreover, they exhibited significant enrichment in metabolically active tissues and specific cell types, further supporting their functional relevance. Collectively, this analytical framework not only deepens our understanding of the genetic architecture underlying ASCVD but also provides a foundational resource for the development of targeted therapeutic strategies.

The ARVCF gene encodes a catenin family member involved in cell-cell adhesion and cytoskeletal organization. While extensively studied in neurodevelopmental contexts, emerging evidence suggests ARVCF may influence vascular wall remodeling and endothelial function, processes fundamental to atherogenesis (Balda and Matter 2003, Huebner et al. 2022). It has been proposed as a modulator of vascular integrity, with implications for plaque stability and inflammatory cascades central to atherosclerotic cardiovascular disease (ASCVD) (Ajoolabady et al. 2024, Ramoni et al. 2025). ARVCF may also contribute to endothelial dysfunction through mechanisms involving oxidative stress (Huebner et al. 2022). Although studies on ARVCF polymorphisms in ASCVD are limited, with no significant associations in large-scale cohorts, our MR analysis indicated a potential causal link between ARVCF expression and ASCVD risk, supported by bioinformatics analyses of adherens junction dynamics. The directionally inconsistent MR estimates for ARVCF across tissues (protective in aorta/coronary but risk‑increasing in tibial) may reflect tissue‑specific regulation, horizontal pleiotropy, or the limitations of single‑instrument MR. Colocalization suggested a shared causal variant in both tissues, but further studies are needed to resolve this discordance. Of note, a prior CRISPRi study targeting the CAD-associated variant at the ARVCF locus (Barbera et al. 2025) showed that ARVCF expression was reduced by approximately 50%, whereas COMT expression was reduced by 80–90%, suggesting that COMT may represent an alternative or additional causal gene at this locus. Therefore, the relative contribution of ARVCF versus COMT to ASCVD susceptibility warrants further investigation. Glutamine-fructose-6-phosphate transaminase 1 (GFPT1), the rate-limiting enzyme of the hexosamine biosynthesis pathway, is a critical nutrient sensor regulating glucose flux into protein O-GlcNAcylation (Sampson et al. 2025). This pathway is fundamental to cellular function, and GFPT1 dysregulation is implicated in metabolic and vascular pathologies (Sampson et al. 2025, Farshadyeganeh et al. 2023). GFPT1 may promote atherosclerosis via O-GlcNAcylation-induced eNOS inhibition and oxidative stress (Sampson et al. 2025). Emerging evidence also links GFPT1 to lipoprotein modifications and hypertriglyceridemia (Farshadyeganeh et al. 2023), impacting insulin resistance and arterial inflammation key to ASCVD (Sampson et al. 2025, Farshadyeganeh et al. 2023). This study used MR analysis to identify a causal link between GFPT1 expression and decreased ASCVD risk, highlighting it as a candidate gene for future functional exploration. Among the tissues examined, colocalization evidence for GFPT1 was strongest in heart atrial appendage (PP.H4 = 0.87), suggesting that this tissue may be more biologically relevant to ASCVD than esophagus muscularis, where colocalization support was only suggestive (PP.H4 = 0.64).

The NFU1 gene encodes a scaffold protein essential for iron-sulfur (Fe-S) cluster biogenesis, primarily within mitochondria (Cai et al. 2016). Mutations in NFU1 are associated with multiple mitochondrial dysfunction syndromes, but no direct link to glycosylation has been established (Camponeschi et al. 2022). Fe-S cluster proteins play a key role in metabolic processes, including oxidative phosphorylation and fatty acid oxidation, suggesting an indirect involvement in myocardial energy metabolism (Read et al. 2021). Mitochondrial disorders such as Friedreich ataxia exhibit both cardiomyopathy and metabolic disturbances (Worth et al. 2015). NFU1 regulates cellular energy homeostasis by supporting the activity of Fe-S-dependent enzymes in the electron transport chain (Cai et al. 2016), implicating NFU1 in ASCVD by modulating mitochondrial function and energy pathways. USP39 (ubiquitin-specific protease 39), a component of the spliceosome complex, regulates pre-mRNA splicing by stabilizing the U4/U5/U6 tri-snRNP complex (Cui et al. 2023). It serves as a potential oncogene in various cancers, and its dysregulation contributes to genomic instability and aberrant cell proliferation (Chen et al. 2025). The role of USP39 in RNA splicing suggests its involvement in vascular endothelial function and potential contributions to ASCVD pathogenesis through altered expression of genes governing lipid metabolism and inflammatory responses (Cui et al. 2023).

The above‑mentioned prioritized genes and their associated pathways have well‑recognized links to immune and inflammatory regulation. Inflammation serves as a critical component of the immune system’s defense mechanism, protecting the host from pathogenic insults and facilitating tissue repair. However, when dysregulated or chronically sustained, inflammatory responses can become maladaptive, contributing to a spectrum of adverse clinical outcomes, including septic shock, hypersensitivity reactions (such as atopy, anaphylaxis, and contact hypersensitivity), transplant rejection, and the pathogenesis of various chronic diseases (Lawrence and Gilroy 2007). The interplay between inflammatory processes and genetic predisposition is increasingly recognized as a key determinant in the initiation and progression of ASCVD (Zhuang et al. 2021). Our group has previously identified metabolic disease loci and estimated the mediating role of inflammatory biomarkers in chronic metabolic disorders (Fu et al. 2025, Fu et al. 2025, Fu et al. 2024, Fu et al. 2025, Fu et al. 2025), further underscoring the relevance of inflammation as a mechanistic link between genetic susceptibility and clinical manifestations of cardiometabolic disease.

Methodological limitations of the statistical approaches used in this study warrant careful consideration. TWAS can produce false positives due to LD contamination or horizontal pleiotropy, and non-colocalized TWAS signals may reflect distinct causal variants rather than genuine gene–trait associations. Moreover, even with COJO and colocalization, TWAS signals may still be influenced by correlated expression of neighboring genes, making it difficult to definitively assign causality to a single gene. Fine‑mapping approaches (e.g., FOCUS or SuSiE‑based TWAS) could further refine causal inference, but these are beyond the scope of the current hypothesis‑generating framework and should be pursued in future studies. Mendelian randomization using a single cis‑eQTL instrument (as was necessary for most genes in this study) is susceptible to weak instrument bias and cannot distinguish horizontal pleiotropy from true causal effects, despite our use of colocalization as a complementary filter. Transcriptome imputation relies on GTEx eQTL models, which may not capture tissue-specific or context-dependent regulation, and the substantial variation in GTEx sample sizes across tissues (73–706) can lead to uneven prediction accuracy. These limitations collectively underscore that our findings should be viewed as hypothesis‑generating, not as mechanistic or therapeutic validation.

This study successfully identified four genes associated with susceptibility to ASCVD—ARVCF (previously implicated in CAD but incompletely characterized in ASCVD), GFPT1, NFU1, and USP39 (newly prioritized in the composite ASCVD phenotype) — and proposed potential roles for these genes in the pathophysiology of these metabolic disorders. By uncovering shared genetic risk factors, this research advances our understanding of the molecular mechanisms linking diverse cardiovascular phenotypes. From a clinical perspective, these findings may inform more precise risk stratification, enabling earlier identification of individuals at elevated risk for ASCVD. Moreover, the identification of genetic markers could guide personalized therapeutic strategies, potentially improving disease management and patient outcomes. From a public health standpoint, integrating genetic insights into population-level screening protocols could facilitate earlier detection of high-risk groups, promoting timely interventions aimed at preventing disease progression. Ultimately, a deeper understanding of the genetic overlap between ASCVD and related traits may inform the development of novel prevention strategies, including gene-based therapies, and contribute to improved health outcomes for affected individuals. Collectively, this research lays the groundwork for future translational efforts aimed at reducing the global burden of ASCVD.

Several limitations of this study warrant consideration. First, potential bias may arise from the age distribution inherent to ASCVD populations, as the condition predominantly affects older adults, potentially introducing age-related confounding. Additionally, collider bias could affect the results if the study populations were not optimally matched, potentially leading to spurious associations. Future studies should address these limitations through stratified analyses, rigorous age adjustment, and methods such as propensity score matching to ensure more robust and generalizable findings. Second, the exclusive focus on individuals of European ancestry limits the generalizability of our results to other populations. Although we restricted all input GWAS to European-ancestry individuals and applied LDSC to correct for sample overlap, residual population substructure or subtle ancestry differences across studies may still affect genetic correlation estimates. This limitation should be considered when interpreting the latent ASCVD factor. Third, despite efforts to minimize false-positive associations through multiple convergent statistical approaches, the absence of an independent replication dataset constrains our ability to fully validate the robustness of the identified genetic signals. To overcome these challenges, future research should prioritize the inclusion of ancestrally diverse cohorts, enabling a more comprehensive characterization of the genetic architecture underlying ASCVD. We acknowledge substantial variation in GTEx v8 tissue sample sizes (73–706), which may affect eQTL model stability and power across tissues. However, the cross-tissue UTMOST framework uses a group-lasso penalty to jointly model eQTL effects across tissues, partially mitigating this imbalance. Additionally, our final gene prioritization relied on convergence across multiple convergent statistical methods and tissues, reducing the risk of tissue-specific power bias. It is important to distinguish between broad expression, tissue enrichment, disease-relevant specificity, and causal tissue assignment. In this study, broad expression alone was not interpreted as evidence of disease relevance; rather, we integrated formal tissue enrichment analysis (TSEA), colocalization, and MR to infer potential tissue-specific roles. The downstream exploratory analyses (e.g., GeneMANIA, pathway enrichment, AlphaFold predictions, CTD interactions, PheWAS) are hypothesis-generating and do not establish mechanistic or therapeutic validity. These analyses are intended to generate biological hypotheses that must be rigorously tested in future experimental studies. Finally, the absence of an independent replication cohort (e.g., FinnGen, separate UK Biobank subsets, or multi-ancestry datasets) is a major limitation. Although we have employed multiple convergent statistical approaches to reduce false positives, replication in independent samples would substantially enhance confidence in the robustness of our prioritized genes and is essential before any translational claims can be made. This study should therefore be viewed as a hypothesis‑generating prioritization framework requiring external validation. Furthermore, our study relies exclusively on GTEx v8 eQTL data derived from non‑diseased donors. The eQTL landscape may differ in patients with established ASCVD. Independent validation using disease‑relevant multi‑tissue expression datasets (e.g., STARNET) or published transcriptomic/proteomic data from human ASCVD samples or animal models would considerably enhance confidence in our prioritized genes. Such validation is beyond the scope of the current hypothesis‑generating framework but should be pursued in future studies. Of note, we did not perform separate TWAS or colocalization analyses for each individual phenotype (e.g., CAD‑only, stroke‑only, PAD‑only) because our primary aim was to prioritize genes underlying shared ASCVD susceptibility. Consequently, whether the identified genes reflect common vascular mechanisms or are driven predominantly by a single trait remains to be determined. Future studies are encouraged to investigate phenotype‑specific contributions to further refine the biological interpretation of these candidate genes.

In summary, our multi-tissue transcriptome-wide association study prioritized novel candidate genes whose expression is associated with susceptibility to ASCVD. This work provides a hypothesis‑generating resource for future investigations aimed at testing the biological significance of these genetic signals. Any deeper understanding of the mechanisms through which these genes might influence disease pathogenesis will require direct experimental validation.

Authors’ contributions

Study concept and design: LF; Acquisition of data: LF; Analysis and interpretation of data: LF and QL; Drafting of the manuscript: LF and QL; Critical revision of the manuscript for important intellectual content: QL, XH, YQH and LF; Funding recipients: LF, QL and YQH.

Funding

This study was supported by grants to L.Q and F.L.W from the Capital’s Funds for Health Improvement and Research (grant no. Capital’s Funds for Health Improvement and Research 2026-2G-2104 and 2024-4-20911), and F.L.W. from the National Natural Science Foundation of China (grant no. 82204063), Young Elite Scientists Sponsorship Program of the Beijing High Innovation Plan (grant no. 20250685), Beijing Chronic Disease Prevention and Health Education Research Association and Zhongguancun Talent Association’s “Future Talents” Training Program in the Medical Engineering Field (grant no. MBRC0012025055), and H.Y.Q from the National Key R&D Program of China (grant no. 2023YFF1205101).

Data availability

All analyses in this study were performed using data that are publicly available or were acquired through institutional applications. Summary-level statistics for coronary heart disease are available at https://ftp.ebi.ac.uk/pub/databases/gwas/summary_statistics/GCST90132001-GCST90133000/GCST90132314; stroke, https://ftp.ebi.ac.uk/pub/databases/gwas/summary_statistics/GCST005001-GCST006000/GCST005838; transient ischemic attack, https://www.finngen.fi/en/access_results; peripheral artery disease, https://ftp.ebi.ac.uk/pub/databases/gwas/summary_statistics/GCST90018001-GCST90019000/GCST90018890; and abdominal aortic aneurysm, https://csg.sph.umich.edu/willer/public/AAAgen2023. Gene expression and eQTL data are freely available at https://ftp.ebi.ac.uk/pub/databases/spot/eQTL/imported/GTEx_V8. GTEx multi-tissue gene expression weight for FUSION analysis (http://gusevlab.org/projects/fusion/#gtex-v8-multi-tissue-expression). LD reference data of 1000 Genomes for FUSION analysis (https://data.broadinstitute.org/alkesgroup/FUSION/LDREF.tar.bz2). Data for UTMOST analysis files (https://zhaocenter.org/UTMOST). Data for MAGMA analysis files (https://fuma.ctglab.nl/downloadPage). All other data supporting the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.