Proteomic signature of metabolic dysfunction-associated steatotic liver disease and risk of atherosclerotic cardiovascular disease

Abstract

Metabolic dysfunction-associated steatotic liver disease (MASLD) is associated with increased atherosclerotic cardiovascular disease (ASCVD) risk. However, evidence on proteomic mechanisms linking MASLD to ASCVD is limited. This study aims to identify proteomic signatures of MASLD and ASCVD subtypes (ischemic heart disease [IHD], peripheral artery disease [PAD], and stroke), evaluate mediating effects of proteins, and develop a proteomic-based ASCVD risk prediction model in MASLD patients. Among 40,913 UK Biobank participants (median follow-up 13.42 years [interquartile range, 12.52–14.22]), 14,425 (35.26%) had MASLD at baseline, and 6,014 (14.70%) developed ASCVD during follow-up (4,420 IHD, 866 PAD, and 1,767 stroke events; subtypes not mutually exclusive). We constructed a binary variable representing proteomics-inferred MASLD (cProMASLD) from MASLD-associated proteins. Two-step Mendelian randomization was applied to assess the mediating effects of proteins associated with MASLD and ASCVD subtypes. Furthermore, we integrated the all shared proteins associated with both MASLD and ASCVD subtypes into the conventional SCORE2 model to develop a prediction model specifically for ASCVD subtypes in the MASLD population, named Pro-SCORE2. Both MASLD and cProMASLD were significantly associated with an increased risk of ASCVD subtypes, with stronger associations observed for cProMASLD (IHD: HR 1.50 [95% CI 1.41–1.60] vs. 1.58 [1.48–1.68]; PAD: 1.25 [1.09–1.44] vs. 1.43 [1.24–1.64]; stroke: 1.19 [1.08–1.31] vs. 1.21[1.10–1.34]). After adjusting for MASLD, cProMASLD remained positively associated with ASCVD risk. This suggests that cProMASLD may capture MASLD-related physiological heterogeneity beyond clinical MASLD classification. We found 15, 3, and 3 proteins mediating the associations of MASLD with IHD, PAD, and stroke, respectively, including FABP4 (MASLD-IHD, mediation proportion: 15.12%), IL7R (MASLD-PAD, 7.45%), and EDA2R (MASLD-stroke, 9.24%). The Pro-SCORE2 significantly improved ASCVD risk prediction in the MASLD population, with a c-index increase of 7.5–9.6% and a 10-year AUC increase of 5.8–9.2% compared to SCORE2. These findings may offer new insights for risk stratification and potential therapeutic targets for ASCVD in MASLD patients.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12933-025-03034-7.

Introduction

Metabolic dysfunction-associated steatotic liver disease (MASLD) is a global public health crisis, affecting approximately 38% of the adult population worldwide [1]. Its clinical harm is not limited to liver complications but is also significantly associated with an increased risk of atherosclerotic cardiovascular disease (ASCVD) [2]. Recent large-scale cohort studies have strengthened this association [3–6]. For example, in a Korean health-screening cohort of 9.8 million adults, MASLD was associated with a 39% higher risk of composite cardiovascular events compared with individuals without MASLD [4]. Similarly, in the multinational prospective cohort including 15,784 participants, MASLD was associated with a 106% higher risk of cardiovascular mortality [5]. International guidelines from organizations such as European Association for the Study of the Liver (EASL) and American Association for the Study of Liver Diseases (AASLD) clearly recommend ASCVD risk assessment for MASLD patients [7, 8]. However, the underlying comorbid mechanisms between MASLD and ASCVD remain poorly understood, highlighting a substantial gap and the lack of precise risk stratification tools based on pathophysiological characteristics.

Recent studies have shown that MASLD and ASCVD share genetic susceptibility loci and environmental exposure profiles [9, 10], suggesting potential common molecular pathways between them. As terminal effect molecules of gene-environment interactions, plasma proteins not only reflect dynamic changes in metabolic homeostasis but also have clinical translational potential as biomarkers [9]. However, previous studies have separately explored the proteomic characteristics of MASLD and ASCVD to identify potential therapeutic targets and high-risk populations [11–15]. Moreover, current research has yet to elucidate the key protein mediators of the progression of MASLD to ASCVD and whether these proteins can contribute to improved ASCVD risk stratification in the MASLD population. Assessing the shared proteomic basis between MASLD and ASCVD is important to better understand potential pathophysiology and for identifying strategies to control ASCVD progression in high-risk MASLD populations [16, 17].

Therefore, this study explores the proteomic associations between MASLD and ASCVD through a multi-layered analysis: first, large-scale proteomic screening will identify shared core proteins between MASLD and ASCVD; secondly, mediation effect models will analyze the mediation proportion of these proteins in the association of MASLD with ASCVD; finally, a protein-based ASCVD risk prediction model for MASLD populations will be developed and validated. This study is the first to integrate full-chain evidence from protein mediator identification to the development of clinical translational tools, offering a novel perspective for the prevention and treatment of MASLD combined with ASCVD.

Materials and methods

Study population

The UK Biobank (UKB) is a large-scale, population-based prospective cohort study that has recruited over 500,000 participants aged 40 to 69 years [18, 19]. This study was approved by the North West Multicenter Research Ethics Committee (11/NW/0382), and informed consent was provided by all participants (application number 98410). For the present analyses, we restricted the sample to participants with plasma proteomics data in UKB (n = 53,014) [20]. Within this proteomics subset, we applied the following exclusions: participants who withdrew consent; those with prevalent ASCVD, hepatic diseases, or liver cancer at baseline, or with liver cancer diagnosed within the first year of follow-up; individuals with alcohol intake > 20 g/day (women) or > 30 g/day (men); those with missing baseline MASLD status; and those who developed ASCVD within the first year of follow-up. After these exclusions, the analytic cohort comprised 40,913 participants (Figure S1).

Assessment of MASLD and ASCVD

MASLD at baseline was defined as the presence of steatotic liver disease (SLD) and one or more of the following cardiometabolic risk factors measured at baseline. In the UKB, due to a lack of liver histology and imaging data, SLD was defined as a fatty liver index (FLI) score ≥ 60. FLI is a composite scoring system that evaluates waist circumference (WC), γ-glutamyltransferase levels, triglyceride levels, and body mass index (BMI) to identify fatty liver disease [21]. The cardiometabolic risk factors for MASLD included: (1) BMI ≥ 25 kg/m² or male waist circumference ≥ 90 cm, female waist circumference ≥ 80 cm; (2) Hemoglobin A1c (HbA1c) ≥ 5.7% (39 mmol/mol), fasting blood glucose ≥ 5.6 mmol/L (≥ 100 mg/dL), 2-hour postprandial blood glucose ≥ 7.8 mmol/L (≥ 140 mg/dL), diagnosed type 2 diabetes, or use of anti-glycemic medications; (3) blood pressure ≥ 130/85 mmHg or use of antihypertensive medications; (4) triglycerides ≥ 1.70 mmol/L (150 mg/dL) or use of lipid-lowering medications; (5) high-density lipoprotein (HDL) cholesterol ≤ 1.0 mmol/L (40 mg/dL) for men, ≤ 1.3 mmol/L (50 mg/dL) for women, or use of lipid-lowering medications [22]. MASLD does not include excessive alcohol consumption or chronic viral hepatitis (HBV or HCV).

ASCVD is defined to include ischemic heart disease (IHD), peripheral artery disease (PAD), and stroke. Incidence was determined based on International Classification of Diseases (ICD-9 and ICD-10) diagnoses during follow-up. The detailed definition of outcomes is presented in Tables S1. All participants were followed from their enrollment date until the outcome of interest, death, loss to follow-up, or the study censoring date (31 October 2022 for England, 31 August 2022 for Scotland, and 31 May 2022 for Wales for hospital inpatient data, and 30 November 2022 for death data), whichever occurred first.

Assessment of covariates

The covariates adjustment set was based on previous literature and included demographic characteristics (age, sex, ethnicity), socioeconomic factors (Townsend deprivation index), lifestyle factors (smoking status, physical activity, dietary patterns), chronic kidney disease, family history of cardiometabolic diseases, and aspirin use [4, 22, 23]. The assessment and categories of covariates are displayed in Table S2. Missing data on covariates were replaced using imputation by chained equations [24]. In our primary analyses, we did not adjust for cardiometabolic risk factors, as they are components of the MASLD diagnostic criteria. In sensitivity analyses, we additionally adjusted for obesity, diabetes, hypertension, and dyslipidemia separately to assess their impact on the results. Furthermore, given that cancer and chemotherapy/radiotherapy may influence proteomic profiles, we excluded participants with cancer at baseline in sensitivity analysis. Detailed definitions for cancer, obesity, diabetes, hypertension, and dyslipidemia are provided in Table S2.

Proteomic profiling

Using the advanced antibody-based Olink Explore 3072 platform, protein expression was standardized and quantified based on baseline blood samples. Detailed information regarding participant selection, Olink proteomics testing, data processing, and quality control has been previously published [25]. A total of 2,923 proteins were measured using this platform. Applying a prespecified exclusion threshold of > 60% missingness, three proteins-GLIPR1 (99.7%), NPM1 (74.0%), and PCOLCE (63.6%)-were removed, leaving 2,920 proteins for analysis (Table S3). A k-nearest neighbors approach (k = 10) was applied to impute missing proteomic data [26]. Protein levels were log2-transformed and standardized before analysis.

Statistical analyses

Figure 1 Overview of this study. Step (1): Identification of MASLD and ASCVD subtype-related proteins using a two-step screening method. Step (2): Construction of a protein risk score based on MASLD-associated proteins and exploration of its association with ASCVD subtypes. Step (3): Utilization of Mendelian randomization to explore the causal relationship between proteins and MASLD/ASCVD. Step (4): Mediation analysis to investigate the mediating effect of proteins in the MASLD-ASCVD association, along with protein-protein interaction network analysis, enrichment analysis, and drug prediction. Step (5): Development of a protein-based ASCVD risk prediction model in the MASLD population.

Fig. 1.

illustrates the overall analytical framework of this study, with detailed descriptions of each step provided below

Identification of core proteins for disease

We used a two-stage modeling strategy to identify plasma proteins significantly associated with MASLD. First, a multivariable logistic regression model was applied to assess the independent associations between each protein and MASLD, adjusting for demographic characteristics (age, sex, ethnicity), socioeconomic factors (Townsend deprivation index), lifestyle factors (smoking status, physical activity, dietary patterns), chronic kidney disease, family history of cardiometabolic diseases, and aspirin use. To account for multiple testing, Bonferroni correction was applied (significance threshold α = 0.05/the total number of proteins tested). Subsequently, proteins found to be statistically significant in the initial analysis were further selected using LASSO regression. To enhance the model’s stability, 100 bootstrap resampling iterations were performed to construct the LASSO model, and only proteins selected in 100% of the iterations were retained as core proteins associated with MASLD. The optimal penalty parameter (λ) was determined through 10-fold cross-validation to balance model complexity and predictive performance.

Similarly, to identify core proteins associated with ASCVD subtypes, a multivariable Cox proportional hazards regression model was applied, adjusting for the same covariates, to assess the association between each protein and the ASCVD subtypes. In addition, a Cox-LASSO regression model was constructed using the same bootstrap stability selection strategy to identify core proteins associated with ASCVD.

Assessment of proteomic MASLD signature

Based on the identified core proteins, we constructed an elastic-net-penalized logistic regression model to MASLD status. The linear predictor from this model was taken as the continuous proteomics-derived MASLD score (ProMASLD):

The ProMASLD score was standardized to have a mean of 0 and a standard deviation of 1 for subsequent analyses. We then applied the inverse-logit to the linear predictor ProMASLD to obtain the predicted probability of MASLD, and dichotomized it at 0.5 to define a binary proteomics-inferred MASLD variable: cProMASLD. Similar weighting methods have been widely used in the calculation of genetic risk scores, metabolomics risk scores, and others [27]. Using Cox proportional hazards models, we evaluated and compared the associations of the clinically defined binary MASLD status (yes/no), cProMASLD (yes/no), continuous ProMASLD (per 1-SD increase), and the traditional fatty liver index (FLI, per 1-SD increase) with each ASCVD subtype after adjustment for covariates. We then fitted joint models including cProMASLD with MASLD and ProMASLD with FLI, respectively, to assess the independent effects of cProMASLD and ProMASLD on ASCVD. A directed acyclic graph (DAG) illustrating the assumed causal relationships among FLI components, estimated MASLD based on FLI, proteins, and ASCVD is provided in Supplementary Figure S2. Additionally, we performed stratified analyses by baseline MASLD status to further evaluate these associations.

Protein-protein interaction and enrichment analysis

We utilized the STRING database to investigate interactions between proteins associated with both MASLD and ASCVD, along with their 10 primary and 10 secondary related proteins (threshold: medium confidence, 0.40). Subsequently, enrichment analysis was performed on the proteins within this interaction network to identify key biological pathways and mechanisms associated with them. The Kyoto Encyclopedia of Genes and Genomes (KEGG) database is a comprehensive and publicly available resource for pathway research. The Gene Ontology (GO) database categorizes the functions of genes and proteins into three categories: biological process, cellular component, and molecular function, facilitating a deeper understanding of their functional roles. Pathways and processes with a false discovery rate (FDR) < 0.05 were considered statistically significant.

Mediation analysis

To further elucidate the mediating role of proteins in the relationship between MASLD and ASCVD, we employed a two-step Mendelian randomization (MR) mediation analysis. This approach, first proposed by Relton and Davey Smith in 2012 as a framework for applying MR in mediation settings, involves the following three two-sample MR analyses: (i) Estimating the total causal effect of MASLD on ASCVD; (ii) Evaluating the causal association between MASLD and proteins; (iii) Assessing the causal association between proteins and ASCVD, using cis protein quantitative trait loci (cis-pQTLs) as instrumental variables for proteins to enhance specificity [28]. All MR analyses rely on the three core assumptions for a valid instrumental variable (IV): relevance (the genetic variant is associated with the exposure), independence (the variant is not associated with confounders), and the exclusion restriction (the variant influences the outcome only through the exposure). In our analysis, strict selection criteria for IVs were applied, ensuring they met the following standards: significant association with the exposure (P < 5e-8), exclusion of linkage disequilibrium (LD, r² < 0.001, kb = 10,000), sufficient association strength (F > 10), and controll for confounding effects using the OpenGWAS database [29]. Proteins without valid IVs were excluded.

We conducted MR analyses with a primary focus on inverse variance weighted (IVW) estimates, which provide unbiased estimates when all IVs are valid. Odds ratios (ORs) with 95% confidence intervals (CIs) were derived from MR analyses and P values were adjusted for multiple testing using the Benjamini-Hochberg (BH) procedure. The mediation proportion is the ratio of the mediation effect to the total effect, with the mediation effect calculated as the product of the mediator-exposure effect and the mediator-outcome effect. To increase the reliability of our results, we complemented this with several supplementary methods: MR Egger regression, which allows detection and correction of directional pleiotropy through its intercept under the instrument strength independent of direct effect (InSIDE) assumption; the weighted median method, which yields consistent estimates if at least 50% of the total weight comes from valid instruments; the weighted mode method, which is valid when a plurality of instruments identify the same causal effect; MR-RAPS to provide more accurate estimates in the presence of weak instruments or horizontal pleiotropy; and MR-PRESSO to identify and correct for outliers caused by horizontal pleiotropy.

The genetic association data used in this study were derived from multiple public databases. The genome-wide association study (GWAS) data for MASLD were sourced from a large-scale study conducted by the UKB, which included 165,984 MASLD cases and 269,322 controls [30]. GWAS data for ASCVD subtypes were obtained from the FinnGen consortium (https://www.finngen.fi/en/access_results), specifically including: IHD dataset (finngen_R12_I9_IHD, including 84,088 cases and 416,260 controls), PAD dataset (finngen_R7_I9_PAD, including 11,924 cases and 288,638 controls), and Stroke dataset (finngen_R12_I9_STR, including 34,110 cases and 450,023 controls). Additionally, pQTL data were sourced from the deCODE genetics study in Iceland, which conducted plasma protein measurements for 4,907 proteins in 35,559 participants [31].

Drug prediction

To further enhance the clinical value of this study, we utilized the Drug-Gene Interaction Database (DGIdb) (https://www.dgidb.org/) to predict potential drugs that may influence the occurrence of ASCVD in MASLD patients [32].

Prediction model construction and evaluation

To evaluate the clinical value of shared proteins between MASLD and ASCVD subtypes in predicting the incidence of ASCVD in MASLD patients, we constructed and compared three prediction models: (i) the SCORE2 model including traditional ASCVD risk factors; (ii) the SCORE2-PRO model integrating all shared proteins between MASLD and ASCVD subtypes in addition to traditional risk factors; and (iii) a proteomics-only model (PRO) comprising the shared proteins without traditional risk factors. We further performed stratification by age (≥ 65 vs. < 65 years), sex, and obesity status to assess the applicability of the model in different subgroups. All prediction models were constructed using the elastic net Cox model. The dataset was randomly split into training and testing sets at a 7:3 ratio. The accuracy of model predictions was assessed using the receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC). Additionally, Harrell’s C-index was used to assess the discriminatory ability of the models before and after adding proteins. To interpret the models and quantify the contribution of each predictor, we computed SHAP (Shapley Additive exPlanations) on the test set, and global importance was summarized by the mean absolute SHAP value [33].

All statistical analyses were performed using R version 4.4.3. All statistical tests were two-tailed, with P-values < 0.05 considered statistically significant.

Results

A total of 40,913 baseline participants were included in this study (median follow-up of 13.42 years [IQR 12.52–14.22]; mean [SD] age 56.54 [8.25] years; 59.43% female), of whom 14,425 (35.26%) had MASLD. The baseline characteristics of the study population, categorized by baseline MASLD status, are shown in Table S4. Generally, participants with MASLD tended to be more socioeconomically disadvantaged, have poorer dietary habits, be less physically active, and had a higher prevalence of hypertension at baseline. During a median follow-up of 13.42 years (interquartile range, 12.52–14.22 years), we observed 4,420 cases of IHD (incidence rate: 8.49 per 1,000 person-years), 866 cases of PAD (1.60 per 1,000 person-years), and 1,767 cases of stroke (3.29 per 1,000 person-years).

Identification of core proteins for MASLD and construction of ProMASLD

We identified 654 core proteins associated with MASLD (Table S5-S6). KEGG analysis showed that MASLD core proteins were predominantly enriched in inflammation- and immune-related pathways, as well as extracellular matrix and adhesion-related pathways (Table S7). GO analysis further indicated that these proteins were mainly localized to the extracellular region or cell surface, participating in immune response and cell signaling, with molecular functions related to receptor binding and ligand activity. Overall, both KEGG and GO results consistently highlight the involvement of these core proteins in inflammatory/immune regulation and extracellular matrix- or cell surface-associated signaling (Table S8-S10).

Based on the identified MASLD core proteins, we derived a continuous proteomics-predicted MASLD score (ProMASLD) and its binary counterpart (cProMASLD). Baseline characteristics stratified by cProMASLD closely resembled those stratified by MASLD status (Table S11). In the overall population, after adjusting for prespecified covariate set defined in the Methods, MASLD was significantly associated with increased risk of ASCVD subtypes: IHD (HR 1.50, 95% CI: 1.41–1.60), PAD (HR 1.25, 95% CI: 1.09–1.44), and stroke (HR 1.19, 95% CI: 1.08–1.31) (Fig. 2a). Notably, the associations of cProMASLD with these ASCVD subtypes were consistently stronger compared to those of MASLD, with HRs of 1.58 (95% CI: 1.48–1.68) for IHD, 1.43 (95% CI: 1.24–1.64) for PAD, and 1.21 (95% CI: 1.10–1.34) for stroke. Furthermore, after additional adjustment for MASLD status, cProMASLD remained significantly associated with increased risks for all three ASCVD subtypes.

Fig. 2.

Independent and joint associations of MASLD, ProMASLD, and FLI with ASCVD subtypes in overall and stratified populations. All models were adjusted for age, sex, race, Townsend Deprivation Index, smoking status, physical activity, diet score, chronic kidney disease, family history of cardiometabolic diseases, and aspirin use

To examine whether baseline MASLD status modified these associations, we further performed stratified analyses according to MASLD status at baseline (Fig. 2b and c). The results showed significant effect modification by MASLD status for the associations of cProMASLD with IHD and PAD (both P for multiplicative interaction < 0.05). Among participants without MASLD at baseline, cProMASLD exhibited stronger associations with higher risks of IHD (HR 1.50, 95% CI 1.31–1.71) and PAD (HR 1.90, 95% CI 1.45–2.49), compared with those among MASLD participants (HR 1.29, 95% CI 1.13–1.46 for IHD, and HR 1.22, 95% CI 0.91–1.65 for PAD). In contrast, for stroke, no significant interaction by MASLD status was observed (P for multiplicative interaction = 0.393).

We also constructed separate and joint models for ProMASLD and FLI in the overall population as well as within MASLD strata. As shown in Fig. 2, both ProMASLD and FLI were significantly associated with increased risks of ASCVD subtypes in the overall population and across MASLD strata. In the joint models, ProMASLD remained significantly associated with all ASCVD subtypes after adjustment for FLI, whereas the associations of FLI were markedly attenuated after adjustment for ProMASLD.

The mediation effect of MASLD on ASCVD via proteins

To better understand the protein mechanisms through which MASLD may influence ASCVD, we initially identified 239, 87, and 99 proteins associated with IHD, PAD, and stroke, respectively (Table S5-S6). Among these, 103, 39, and 46 proteins also appeared in the MASLD-related protein list (Table S12). These overlaps suggest that MASLD and ASCVD subtypes share common protein mechanisms, providing a foundation for further investigation of their comorbid relationship. Therefore, we conducted MR analysis to assess the mediation effects of these proteins between MASLD and ASCVD subtypes.

The MR analysis results indicated a significant causal relationship between MASLD and ASCVD subtypes. Specifically, using IVW method, the ORs (95% CIs) were as follows: for IHD, 1.21 (1.12–1.31); for PAD, 1.32 (1.19–1.48); and for stroke, 1.07 (1.01–1.13) (Fig. 3a, Table S13). Subsequently, we conducted two-step MR mediation analyses to assess the mediating effects of 103, 39, and 46 proteins identified as intersecting proteins associated with MASLD and IHD, PAD, and stroke, respectively. Of these, 68, 30, and 31 proteins were excluded from further analysis due to insufficient IVs in either the MASLD-to-protein or protein-to-ASCVD subtype analyses. After BH correction, 15, 3, and 3 proteins were found to significantly mediate the associations of MASLD with IHD, PAD, and stroke, respectively (Fig. 3b; Table S14-S18). Among these proteins, 12 exhibited positive mediation effects, while 7 showed negative mediation effects. Notably, FABP4, IL7R, and EDA2R exhibited the highest mediation proportions in the associations between MASLD and each ASCVD subtype, with FABP4 mediating 15.12% of the association with IHD, IL7R mediating 7.45% of the association with PAD, and EDA2R mediating 9.24% of the association with stroke. In comparison, the mediation proportions of the other proteins were significantly lower.

Fig. 3.

Mediation analysis of the association between MASLD and ASCVD subtypes through proteins using a two-step MR mediation method. Panel (a) displays the total effect of MASLD on ASCVD, while Panel (b) shows the MASLD-protein association, the protein-ASCVD association, and the mediation proportion

Protein network construction, functional enrichment, and drug prediction

Protein-protein interaction analysis identified 10 primary and 10 secondary associated proteins related to the 19 proteins that mediate the relationship between MASLD and ASCVD, constructing an interaction network containing 180 protein interactions (Fig. 4a, Table S19). Enrichment analysis revealed that these proteins are involved in 27 biological processes, 8 cellular components, and 9 molecular functions, and are significantly enriched in 5 KEGG pathways (Fig. 4b, Table S20-S23). GO analysis indicated that the proteins were primarily enriched in the extracellular region and cell periphery, participating in biological processes such as ceramide metabolism, sphingolipid metabolism, and glycosphingolipid metabolism, which substantially overlap with the KEGG sphingolipid metabolism and signaling pathways. Their molecular functions included hydrolase activities (e.g., glycosidic bond hydrolases), corresponding to the KEGG lysosome and other glycan degradation pathways. Moreover, the KEGG “metabolic pathways” overlapped broadly with GO lipid metabolic processes, underscoring the coordinated role of these proteins in metabolic regulation.

Fig. 4.

Proteins associated with MASLD and ASCVD: interaction network analysis, KEGG and GO enriched pathways, and potential drug predictions. GO enrichment results show the top five pathways in each category, and drug predictions list the top 10 potential drugs for proteins with positive mediation effects, ranked by interaction score

Using the DGIdb database, we performed drug prediction for 13 of the 19 proteins with positive mediation effects, ultimately identifying 49 potential drugs (Fig. 4c, Table S24). Among these, 20 drugs have been reported to be associated with diseases such as coronary heart disease, myocarditis, hypertension, and diabetes. For instance, Remikiren and VTP-27,999, two investigational antihypertensive drugs targeting REN, both had interaction scores of 15.66, underscoring their established role in hypertension therapy. Additionally, we identified 28 previously unreported drugs, opening new potential therapeutic avenues, with (S)-RP-6306 showing the highest predicted score (26.10), followed by ODE-127 (17.40) and Remikiren (15.66).

Prediction model performance

The SCORE2-PRO model, an extension of SCORE2 that integrates shared proteins between MASLD and ASCVD subtypes (n = 103, 39, and 46), showed superior predictive performance over the traditional SCORE2 model. During the 15-year follow-up period, the SCORE2-PRO model outperformed the SCORE2 model in predicting short and long term risks for IHD, PAD, and Stroke (Fig. 5, Table S25-S27). In 10-year risk prediction for IHD, PAD, and stroke, the AUCs of SCORE2 were 0.69 (0.65–0.72), 0.76 (0.69–0.82), and 0.68 (0.62–0.74), while the SCORE2-PRO model improved AUC to 0.73 (0.70–0.77), 0.83 (0.77–0.89), and 0.74 (0.68–0.79), an increase of 5.8–9.2%. The Harrell’s C-index for SCORE2-PRO was significantly higher than SCORE2, with relative increases of 7.5–9.6% (IHD: 0.72 vs. 0.67; PAD: 0.80 vs. 0.73; stroke: 0.73 vs. 0.67), indicating improved discrimination of high-risk groups. We also constructed a proteomics-only model (PRO) using all shared proteins between MASLD and ASCVD. The PRO model performed comparably to SCORE2-PRO across ASCVD subtypes, with similar AUCs and C-index, highlighting the strong prognostic value of these proteins. To further examine whether the predictive information carried by proteins was independent of traditional risk factors, we reconstructed the prediction models using protein residuals after removing the influence of SCORE2 traditional risk factors. Across subtypes, the C-index decreased after residualization (Tables S28-S29).

Fig. 5.

Performance of ASCVD risk prediction models in the MASLD population. Time-dependent area under the curve (AUC) and Harrell’s concordance index (C-index) were used to evaluate and compare the performance of SCORE2, SCORE2-PRO, and PRO models. The top 15 variables contributing most to prediction in SCORE2-PRO are visualized using SHAP

Feature importance analysis based on SHAP value revealed the key role of MASLD-ASCVD shared proteins in risk prediction. As shown in Fig. 5, in the IHD prediction model, the top two features by SHAP were age (0.0088) and CTHRC1 (0.0061). For PAD, CXCL17 (0.0053) was the strongest predictor, followed by type 2 diabetes (0.0039) and smoking (0.0034). For stroke, age (0.0077), NEFL (0.0075), and ITGAV (0.0062) had the highest SHAP value (Tables S30-S32). Several proteins with high SHAP importance also exhibited large coefficient changes after residualization, including CTHRC1 in the IHD model, CXCL17 in the PAD model, and NEFL in the stroke model (Table S33).

Sensitivity analysis

The identification of MASLD-related proteins, ASCVD-related proteins, and their overlap were not substantially altered in sensitivity analyses that (i) excluded participants with any cancer at baseline and (ii) additionally adjusted, in separate models, for hypertension, diabetes, dyslipidemia, or obesity (Table S34-S36). The reconstructed ProMASLD-ASCVD association likewise remained consistent in both direction and magnitude (Table S37). The sensitivity MR results were consistent across MR Egger, weighted median, weighted mode, and MR-PRESSO methods, and pleiotropy testing of the IVs indicated that the exposure effects on outcome were not mediated through non-exposure pathways (horizontal pleiotropy test P > 0.05, Table S3, S15-S18). Stratified analyses by age (≥ 65 vs. < 65 years), sex, and obesity status showed that the proteomics-based prediction model consistently outperformed the traditional SCORE2 model across all subgroups (Table S38-S39).

Discussion

In this large-scale proteomic analysis, the protein risk score constructed based on MASLD-related proteins demonstrated a stronger association with ASCVD subtypes compared to the binary MASLD classification. Mediation analysis identified 15, 3, and 3 proteins as potential mediators in the association of MASLD with IHD, PAD, and stroke, respectively. Notably, FABP4, IL7R, and EDA2R mediated the largest proportion of the associations for IHD, PAD, and stroke, respectively. These mediating proteins are primarily involved in key biological pathways related to lipid metabolism, cytokine regulation, and immune response. Furthermore, incorporating shared proteins associated with both MASLD and ASCVD subtypes into the traditional SCORE2 model significantly improved the predictive accuracy for ASCVD subtypes in the MASLD population.

Several studies have analyzed the plasma proteome of MASLD patients. Liu et al. recently integrated proteins related to liver steatosis using the UKB and Mendelian randomization, identifying several candidate pathogenic proteins [43]. Sveinbjornsson et al. conducted a comprehensive plasma protein analysis in 35,559 Icelandic and 47,151 UKB participants, identifying proteins involved in MASLD pathogenesis [13]. Wu et al. also screened MASLD-associated proteins using UKB [34]. In our study, 341 of 654 MASLD-associated proteins overlapped with those reported by Wu et al., while the remaining differences likely reflect discrepancies in cohort inclusion/exclusion criteria and covariate adjustment strategies. Based on the identified core proteins, we constructed a proteomics-derived MASLD score, represented as both a dichotomized variable (cProMASLD) and a continuous variable (ProMASLD). cProMASLD showed stronger associations with ASCVD risk than the clinical MASLD diagnosis. Notably, the associations of cProMASLD with increased risks of IHD and PAD were stronger among participants without MASLD. We speculate that this population’s lower baseline risk allows cProMASLD to more sensitively capture subclinical atherosclerosis and lipid disturbances related to IHD and PAD, resulting in higher relative risks [35]. By contrast, the higher baseline risk in MASLD participants attenuates the relative effect (HR) of a given absolute increase. The opposite pattern observed for stroke may be attributed to its heterogeneous etiology, including hypertension, cerebral small vessel disease, and atrial fibrillation, therefore, cProMASLD may capture signals of systemic inflammation and prothrombotic tendency more prevalent in MASLD [36].

Furthermore, we examined the independent and joint effects of FLI and ProMASLD on ASCVD in the overall population, as well as among participants with and without MASLD. Because MASLD in this study was defined as an estimated status based on dichotomized FLI and cardiometabolic risk factors, this definition may not fully capture the associations of FLI—or the true MASLD state—with ASCVD. Consequently, a positive association between FLI and ASCVD remained among participants without MASLD. In joint models including both ProMASLD and FLI, proteomic signatures partially mediated the association between FLI and ASCVD. Therefore, the association of FLI with ASCVD was attenuated after adjustment for ProMASLD. However, ProMASLD remained significantly associated with ASCVD after adjusting for FLI, suggesting that it captures molecular signals beyond obesity and lipid burden—such as inflammation (e.g. CXCL9 and TNFSF10), lipoprotein metabolism (e.g. APOM and PON1), and oxidative stress (e.g. SELENOP and EDN1)—that more directly reflect subclinical pathology [37–40]. These findings are based on relative risk estimates and may be affected by limited power in stratified analyses and residual confounding; the underlying mechanisms warrant further investigation. Overall, the proteomics-derived MASLD score provides independent, early risk information beyond traditional measures of MASLD and may complement existing tools to improve risk stratification.

Our two-step MR mediation analysis offers the first causal evidence that proteins mediate the link between MASLD and ASCVD. We identified 15, 3, and 3 proteins potentially mediate the relationships of MASLD with IHD, PAD, and Stroke, respectively. Notably, FABP4, a lipid-binding adipokine secreted by adipose tissue and macrophages, reinforced our findings. FABP4 is a marker of hepatic steatosis in MASLD and is linked to disease progression via inflammation and fibrosis [41]. Elevated FABP4 levels have been associated with IHD and shown to accelerate atherosclerosis [42]. By promoting insulin resistance, inflammation, and lipotoxicity, FABP4 disrupts MASLD metabolism and creates a pro-atherogenic environment, serving as a key mediator between liver and cardiovascular disease [43]. In contrast, IL7R and EDA2R, mediating MASLD-PAD/stroke, lack direct cardiovascular evidence but have preliminary support for roles in specific disease processes. IL7R may drive atherosclerosis via immune-inflammatory regulation in adipose tissue [44], while EDA2R, an aging-related gene, may link metabolic abnormalities to stroke through inflammation triggered by its ligand upregulation [45]. Genetic data strengthen their clinical value and highlight them as potential therapeutic targets to reduce ASCVD risk in MASLD. Nevertheless, the specific mechanisms of these proteins in cardiovascular diseases require further investigation. We also observed proteins with negative mediation proportions. For example, MR showed a positive association between MASLD and MMP12; experimental evidence suggests MMP12 upregulation during liver injury may promote tissue repair [46]. However, MR indicated that MMP12 was inversely associated with PAD and stroke, consistent with previous protective associations [47, 48]. We hypothesize that in the early stages of atherosclerosis, MMP12 upregulation may counteract pro-atherogenic activity through extracellular matrix remodeling or inflammation modulation. These findings suggest proteins such as MMP12 may play a causal role in MASLD progression to ASCVD, though their precise mechanisms require further investigation.

In terms of potential interventions, we identified (S)-RP-6306 from the DGIdb database, which showed a high predicted interaction with FABP4 and is an oral small-molecule PKMYT1 kinase inhibitor currently in clinical trials for solid tumors [49]. Similarly, IL7R is predicted to interact strongly with OSE-127, a monoclonal antibody targeting IL-7Rα for inflammatory and autoimmune diseases [50]. Reducing FABP4 activity may improve insulin sensitivity, liver steatosis, and atherosclerosis-related inflammation, while targeting cytokine pathways such as OSE-127 may modulate immune responses; thus, these agents may represent promising strategies to reduce cardiovascular risk in MASLD [51]. As these findings are based on MR inference and computational predictions without clinical validation, their relevance to metabolic disease treatment requires further mechanistic and clinical studies. We also identified Remikiren and VTP-27,999 as REN-targeting candidate drugs with potential relevance to cardiovascular therapy. Previous studies have shown that the metabolic dysregulation and inflammation of MASLD can activate the renin-angiotensin system (RAS), thereby inducing oxidative stress, endothelial dysfunction, inflammation, and fibrosis—all key processes in the initiation and progression of atherosclerosis [52]. RAS blockade has been reported to improve cardiovascular outcomes and, in some cases, to ameliorate hepatic steatosis and fibrosis [53, 54]. However, Remikiren and VTP-27,999, as direct renin inhibitors, remain unapproved, with existing studies focusing mainly on cardiac contractility and hypertension rather than MASLD or IHD [55]. Thus, REN may represent a critical molecular node mediating the MASLD-IHD relationship, but its clinical translation requires validation in MASLD populations.

Our MASLD-specific risk model, integrating shared MASLD-ASCVD proteins, showed superior predictive performance compared with SCORE2, underscoring the importance of tailoring risk prediction tools for high-risk subgroups. MASLD patients are known to have a higher risk of ASCVD, but general risk calculators like SCORE2 often underestimate their true risk because they fail to capture the metabolic and inflammatory abnormalities characteristic of MASLD [56]. Incorporating shared MASLD-ASCVD proteins into the model addressed this limitation. Furthermore, the PRO model performed comparably to SCORE2-PRO and even slightly outperformed it in some cases. This may be explained by two main factors. First, many proteins integrate the downstream effects of traditional risk factors through inflammatory and metabolic pathways, resulting in overlapping information. In penalized regression, when correlated features are present, the model tends to retain the more informative variables while shrinking others, thereby reducing overall information efficiency. Second, traditional risk factors are often measured coarsely, potentially introducing noise and modestly lowering predictive performance in certain settings.

To further investigate the relationship between proteins and traditional risk factors, we reconstructed the models after removing the variance in proteins explained by SCORE2 variables and observed a decline in predictive performance. This finding suggests that several proteins strongly influenced by traditional risk factors and contributing substantially to prediction may partly mediate the effects of these factors on ASCVD outcomes. For example, CTHRC1 may mediate vascular remodeling induced by hypertension, diabetes, and dyslipidemia, contributing to plaque progression and instability [57]; CXCL17 may mediate the link between diabetes- and metabolism-induced inflammatory responses and vascular inflammation [58]; and NEFL may mediate neurovascular injury caused by metabolic risk factors such as diabetes, contributing to cerebrovascular events [59]. In summary, our findings demonstrate that the proteomics-based MASLD-specific risk model can improve ASCVD risk stratification in high-risk populations. Future studies should focus on validating this model in diverse populations and confirming its broader clinical value. In addition, applying feature-decorrelation or ensemble modeling strategies may help mitigate the effects of feature collinearity and enhance model robustness, while integrating additional clinical variables or multi-omic biomarkers (e.g., genomics, metabolomics) could further improve predictive accuracy by capturing the complex interactions between MASLD and ASCVD.

Our study offers new insights into the limited evidence regarding the relationship between proteomics, MASLD, and ASCVD risk. Moreover, as one of the largest cohorts currently encompassing extensive Olink proteomics data, the UKB provides a broad range of protein measurements, helping to explore various biological pathways associated with proteins. Furthermore, the large sample size and long follow-up duration provided by the UKB strengthen the robustness of our results, ensuring more accurate estimates and reliable inferences. However, our study has some limitations. First, despite controlling for many relevant confounders, residual confounding cannot be completely excluded. Nevertheless, when exploring the association between MASLD and ASCVD, we have corrected for key confounding factors. Second, our mediation analysis assumes temporal sequencing between exposure, mediator, and outcome variables. However, proteins and MASLD in our study were assessed at baseline. In addition, the selection of proteins in this study was based on clinically defined MASLD (FLI ≥ 60 combined with at least one cardiometabolic risk factor), rather than imaging- or biopsy-based quantification of hepatic steatosis. Therefore, some degree of misclassification may exist, and this definition may not fully capture the histological heterogeneity of liver pathology. Meanwhile, ProMASLD represents proteomic signals associated with the estimated MASLD status rather than directly measured hepatic pathology, and thus may still deviate from the true underlying MASLD signal. Finally, since the UKB participants are primarily of European ancestry, it is necessary to validate the generalizability and reproducibility of our results in other more diverse and independent populations.

Conclusion

In conclusion, this study reveals the shared proteins and their functional pathways between MASLD and ASCVD subtypes, including IHD, PAD, and stroke. The potential mediator identified, such as FABP4, IL7R, and EDA2R, may offer new directions for targeted therapy. By developing a ASCVD risk prediction model in MASLD participants, we further address the limitations of traditional risk assessment tools, providing new insights for ASCVD risk stratification and intervention strategies in this population. Future studies should further explore the mechanistic roles of these key proteins and validate the clinical applicability of the newly proposed prediction models and therapeutic targets, ultimately aiming to improve clinical outcomes for MASLD and ASCVD patients.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

LLP conceived the idea and contributed to statistical analysis, interpretation of data, and the draft of the manuscript. MJS contributed to the statistical methods and interpretation of data. YHL, CH, and RLL contributed to data verification and manuscript revision. GYQ and YFY contributed to the conception of the study, overall supervision, and final editing of the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by National Natural Science Foundation of China (No. 82473724 to GQ, 82273730 to YY), Shanghai Municipal Natural Science Foundation (22ZR1414900 to YY), the Three-Year Public Health Action Plan of Shanghai (GWVI-11.2-XD10 to YY), and Shanghai Municipal Science and Technology Major Project (ZD2021CY001 to GQ). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This study was approved by the North West Multicenter Research Ethics Committee (11/NW/0382), and informed consent was provided by all participants (application number 98410).

Competing interests

The authors declare no competing interests.