Quantifying the Independent and Combined Contributions of Clinical, Genetic, Proteomic, and Metabolomic Signals to ASCVD Prediction

Abstract

Atherosclerotic cardiovascular disease (ASCVD) risk prediction based on clinical factors and polygenic risk scores (PRS) does not fully capture dynamic biological processes related to near-term risk. Using UK Biobank participants with matched genetics, plasma proteomics, metabolomics, and clinical data (n=15,787; 1,731 incident ASCVD events), we evaluated the independent and combined contributions of clinical, genetic, proteomic, and metabolomic signals within a unified modeling framework. Modality-specific risk scores were constructed for PRS, proteomics (ProRS), metabolomics (MetRS), and clinical factors, then integrated using Cox models. Proteomics provided the strongest independent predictive contribution, PRS contributed orthogonal inherited risk information, and metabolomics provided intermediate prognostic value. Multi-modal integration improved discrimination compared with single-modality models, and joint ProRS-PRS stratification identified biologically and clinically distinct risk subgroups. These findings support multi-modal risk modeling, particularly proteomics-integrated approaches, for improving ASCVD risk stratification and precision prevention.

1. Introduction

Atherosclerotic cardiovascular disease (ASCVD) is a major cause of morbidity and mortality worldwide and remains a central target for prevention efforts. Although clinical guidelines rely heavily on traditional cardiometabolic risk factors to estimate future events, a substantial proportion of individuals who experience ASCVD lack markedly elevated clinical risk at baseline. This gap highlights the need for improved approaches that capture earlier and more nuanced biological signals of disease susceptibility. ASCVD develops through the interplay of inherited susceptibility and dynamic physiological processes that evolve across the lifespan. Polygenic risk scores (PRS), which aggregate the effects of numerous common genetic variants, provide a stable measure of lifelong genetic liability and have demonstrated predictive value across multiple cardiovascular outcomes^{1, 2}. However, genetic predisposition alone does not fully capture ASCVD vulnerability³. Many proximal determinants of risk, such as inflammation, metabolic dysregulation, endothelial dysfunction, and cardiometabolic stress, consider an individual’s current physiological state rather than inherited genomic variation. As a result, PRS offer an incomplete representation of the biological processes driving near- and intermediate-term ASCVD events.

High-throughput molecular profiling technologies, now available in large population biobanks such as UK Biobank (UKB), enable direct measurement of circulating biomarkers that capture individual’s current health status⁴. Plasma proteomics provides diverse pathways including immune activation, vascular remodeling, and tissue injury, whereas metabolomics represents metabolic flux, lipid handling, and energy balance^{5, 6}. These dynamic molecular signatures can vary by age, environment, lifestyle, and subclinical disease progression, and thus may reveal components of ASCVD risk not reflected in genetics or conventional clinical measurements. Clinical risk factors themselves provide another complementary axis, summarizing traditional cardiometabolic traits such as blood pressure, glycemic control, and renal function that remain central to ASCVD pathobiology^{7, 8}.

Despite increasing availability of multi-omics resources, few studies have directly compared the predictive value of clinical variables, PRS, proteomics, and metabolomics within the same cohort⁹. Existing analyses typically evaluate modalities in isolation, limiting the ability to quantify their shared versus distinct contributions or to determine how best to integrate them for risk estimation¹⁰. In addition, although proteomics and metabolomics have shown promise individually, their incremental and joint utility relative to genetics and clinical risk factors remains insufficiently defined¹¹. A unified analytic framework is needed to evaluate these modalities side-by-side, quantify their independent prognostic contributions, and assess whether combining them provides meaningful gains in ASCVD prediction.

Utilizing multi-modal data from UKB, we systematically evaluated four modalities for ASCVD prediction: PRS, plasma proteomics, plasma metabolomics, and clinical risk factors. We compared predictive performance across modalities, examined their correlation structure to characterize biological dependencies, and assessed their combined prognostic value through integrated Cox modeling. We further investigated joint stratification using two modalities to evaluate how inherited susceptibility and current molecular signals interact to shape ASCVD risk. By characterizing the complementary and distinct contributions of genetic, molecular, and clinical information, this study establishes an informatics-driven framework for multi-modal ASCVD risk prediction and clarifies how each data layer contributes to population-scale cardiovascular risk assessment.

2. Methods

2.1. Study design and cohort description

This study adopted a prospective cohort design using the UKB. Baseline was defined as the date of enrollment corresponding to the blood collection visit. To construct an incident ASCVD cohort, we excluded all participants with any ASCVD diagnosis prior to baseline and individuals with documented congenital heart disease^{12, 13}. The analytic population therefore consisted of participants who were ASCVD-free at baseline and were followed longitudinally for new-onset ASCVD, while those who remained event-free throughout follow-up served as controls. Multi-modal omics data were obtained from UKB, which includes genome-wide genetic data, plasma proteomics, metabolomics, and linked electronic health records. Analyses were restricted to individuals with concurrent availability of all three molecular modalities (genetics, proteomics, and metabolomics) and complete baseline clinical covariates relevant to cardiometabolic risk assessment¹⁴. After applying all inclusion and quality-control criteria, the final cohort consisted of 1,731 participants with incident ASCVD and 14,056 controls. The full analytic workflow is summarized in Figure 1A.

Figure 1.

Integrated study design and characteristics of the analytic cohort. (A) Schematic overview of the construction of modality-specific ASCVD risk scores. Genetic, clinical, proteomic, and metabolomic features were each used to derive independent risk scores, which were subsequently integrated within a Cox proportional hazards framework to model incident ASCVD risk. (B) Distribution of follow-up time among 15,787 participants included in the analytic cohort. Median follow-up was 7.8 years for cases and 11.1 years for censored individuals. (C) Baseline characteristics of individuals who developed incident ASCVD compared with event-free controls.

2.2. ASCVD outcome definition and follow-up

ASCVD events were defined using linked UKB hospital inpatient records and standardized ICD-9/10 diagnostic codes. Incident ASCVD encompassed the first occurrence of coronary artery disease (ICD-9 410-414; ICD-10 I21-I25), peripheral arterial disease (ICD-9 440-443; ICD-10 I70-I73), or ischemic stroke (ICD-9 434, 436; ICD-10 I63-I64)¹⁵. Participants were followed from baseline until the earliest of incident ASCVD, death, loss to follow-up, or administrative censoring. Among the 15,787 individuals in the analytic cohort, 1,731 experienced an incident ASCVD event. Median follow-up was 7.8 years among cases and 11.1 years among censored individuals, corresponding to an incidence density of 11.2 events per 1,000 person-years (Figure 1B). Baseline demographic and clinical characteristics of cases and controls are shown in Figure 1C.

2.3. Modality-specific risk score construction

Four modality-specific risk scores were constructed to capture the contributions of genetics, plasma proteomics, plasma metabolomics, and clinical factors to ASCVD risk. The PRS for coronary artery disease was obtained from the UKB pre-computed PRS resource (data-field 26227), standardized to zero mean and unit variance, and used directly without recalibration¹⁶. For the proteomic risk score (ProRS), plasma protein abundances quantified by the Olink Explore 3072 platform were processed using a previously described UKB-PPP quality-control workflow, retaining 2,920 proteins after QC and 10-nearest neighbor imputation of missing NPX values^{5, 10, 17}. Preprocessing was performed in a disease-agnostic manner without using ASCVD outcome labels. An Elastic Net logistic regression model was trained on all proteins in the training subset, with hyperparameters selected through grid search with 10-fold cross-validation. Predicted probabilities served as the ProRS. For the metabolomic risk score (MetRS), biomarkers from the Nightingale NMR platform underwent quality control and preprocessing following previously described UKB metabolomics procedures⁶. In the present study, median imputation and winsorization were performed as disease-agnostic preprocessing steps without using ASCVD outcome labels. For model fitting, metabolomic features were subsequently standardized using parameters estimated in the training subset and applied to the held-out test subset. An Elastic Net logistic regression model with cross-validated hyperparameters was applied, and the resulting predicted probabilities formed the MetRS. The clinical risk score (Clinical RS) was derived from multivariable logistic regression using variables included in the PREVENT equation: age, sex, systolic blood pressure, antihypertensive medication, lipid-lowering medication, smoking status, diabetes status, BMI, HDL, TC, and eGFR¹⁴. All models were trained using 70 percent of the cohort and evaluated in the held-out 30 percent using stratified

sampling to preserve event proportions. Each modality-specific score was standardized before integration into the multi-modal Cox model. The modality-specific logistic models were used to construct parsimonious risk scores that summarize cross-sectional risk information within each data modality and provide a common scale for downstream integration. Prognostic performance under censoring was subsequently assessed using survival-based analyses (e.g., Cox models, Kaplan-Meier curves, and time-dependent AUC), in which these modality-specific scores were treated as predictors rather than as direct estimators of absolute time-to-event risk.

2.4. Integrated multi-modal Cox modeling

To quantify the independent and combined prognostic value of the four modality-specific risk scores, the standardized PRS, ProRS, MetRS, and Clinical RS were entered jointly into a Cox proportional hazards model fit in the training subset, and prognostic performance was evaluated in the held-out test subset. This workflow preserved separation between model development and evaluation by deriving modality-specific scores in the training subset and applying the trained models to the held-out test subset. Hazard ratios were estimated for each modality, and proportionality assumptions were tested using Schoenfeld residuals. The relative prognostic contribution of each modality was assessed through nested-model comparisons, where the decrement in log-likelihood after removing each modality reflected its incremental value. Discriminative performance for individual modalities and their combinations was evaluated using the concordance index (C-index) and Uno’s time-dependent AUC from 1 to 10 years.

Risk stratification analyses were performed by dividing participants into quantiles of each risk score. Kaplan-Meier curves were used to characterize separation in ASCVD incidence across strata. Joint stratification (e.g., PRS × ProRS) was used to evaluate complementarity between modalities and to assess whether molecular profiles enhanced risk separation beyond clinical measures alone.

2.5. Evaluation metrics and risk stratification procedures

Time-dependent discrimination was evaluated using Uno’s estimator of the cumulative AUC, which accounts for right censoring through inverse probability of censoring weights (IPCW)^{18, 19}. IPCW weights were estimated using Kaplan-Meier estimates of the censoring distribution within the held-out test set. All AUC calculations used the full follow-up time without artificial truncation. The cumulative AUC was computed annually from 1 to 10 years after baseline to quantify temporal changes in predictive accuracy across modalities.

Risk stratification analyses were conducted using quantile-based grouping of each modality-specific risk score. For PRS and ProRS, participants were categorized into tertiles (low, medium, high) based on the empirical distribution within the test set, and the same cut points were applied consistently for joint stratification analyses to ensure reproducibility. Kaplan-Meier estimators were used to assess differences in ASCVD incidence across strata, with participants censored at death, loss to follow-up, or end of administrative follow-up. Ten-year ASCVD risk estimates were derived directly from the Kaplan-Meier survival function as 1 − S(t) at t = 10 years. In joint stratification (e.g., PRS × ProRS), survival curves were plotted for all nine combinations to evaluate whether integrated molecular information improved separation of absolute event rates beyond single-modality stratification.

2.6. Statistical Analysis

Analyses were performed using Python (version 3.11) with scikit-learn, lifelines, and statsmodels. Continuous variables were summarized as mean ± SD and categorical variables as proportions. Logistic regression-based discrimination was evaluated using AUC, and longitudinal discrimination was quantified using Uno’s time-dependent AUC with inverse probability of censoring weights. Cox models for integrated risk-score analyses were fit in the training subset, and predictive performance was evaluated in the held-out test subset to preserve separation between model development and evaluation. Incremental prognostic value of each modality was evaluated through changes in log-likelihood, C-index, and stratified Kaplan-Meier analyses. Two-sided P < 0.05 was considered significant for Cox regression, with an emphasis on effect sizes and predictive accuracy rather than hypothesis testing.

3. Results

3.1. Baseline Characteristics

Among the 15,787 participants included in the analytic cohort, 1,731 developed incident ASCVD during follow-up and 14,056 remained event-free. Baseline characteristics are summarized in Figure 1C. Compared with controls, individuals who developed ASCVD were older (60.5 ± 6.9 vs. 56.5 ± 8.1 years) and included a higher proportion of males (59.6% vs. 44.0%). They also had higher systolic blood pressure (144.8 ± 18.5 vs. 137.6 ± 18.4 mmHg), greater use of antihypertensive medication (30.4% vs. 15.5%) and lipid-lowering medication (25.7% vs. 11.7%), a higher prevalence of diabetes (10.7% vs. 4.1%), and higher BMI (28.4 ± 4.9 vs. 27.4 ± 4.7 kg/m²). HDL cholesterol levels were lower among cases (52.0 ± 14.0 vs. 56.2 ± 14.3 mg/dL), as was total cholesterol (218.8 ± 46.7 vs. 223.2 ± 43.7 mg/dL). Kidney function was modestly reduced in ASCVD cases, reflected by lower eGFR at baseline (83.3 ± 18.5 vs. 89.3 ± 17.2 mL/min/1.73 m²).

3.2. Modality-Specific Risk Scoring Model Construction

Elastic Net logistic regression models were trained in the 70 percent training subset to predict incident ASCVD. For the proteomics model, 117 of 2,920 proteins had non-zero coefficients and were retained in the ProRS. The metabolomics model identified 92 informative biomarkers that comprised the MetRS. Hyperparameters for both models were selected using grid search with 10-fold cross-validation, and predicted probabilities were computed in the held-out test set. The Clinical RS was derived using multivariable logistic regression applied to the 11 PREVENT covariates. Coefficient directions were consistent with established ASCVD risk patterns, with positive contributions from age, male sex, systolic blood pressure, antihypertensive therapy, lipid-lowering therapy, smoking, diabetes, and BMI, and inverse associations for HDL and eGFR, whereas total cholesterol showed minimal influence, potentially reflecting overlap with other clinical covariates and the inclusion of lipid-lowering medication use in the model.

The four modality-specific models showed distinct predictive performance for incident ASCVD in the held-out test set. In cross-sectional discrimination, ProRS achieved the highest AUC (0.702), followed by MetRS (0.625), the Clinical RS (0.696), and the PRS (0.577) (Figure 2A). These results indicate that plasma proteomics provided stronger predictive information than clinical variables or genetics alone. Longitudinal analyses using time-dependent AUC from 1 to 10 years demonstrated consistent performance rankings across follow-up (Figure 2B). ProRS maintained the highest discriminative ability over time (AUC range: 0.713-0.761), whereas the Clinical RS showed stable but moderately lower performance (AUC range: 0.702-0.725). MetRS exhibited intermediate performance (AUC range: 0.632-0.696), while PRS provided the weakest temporal discrimination throughout follow-up (AUC range: 0.584-0.648). Risk score distributions differed between cases and controls for all four modalities (Figure 2C), with ASCVD cases exhibiting consistently higher scores than controls. The separation was most pronounced for ProRS and the Clinical RS, modest for MetRS, and smallest for PRS. These findings demonstrate that proteomic and metabolomic signals contribute meaningful predictive information beyond traditional clinical factors and genetic susceptibility.

Figure 2.

Comparative performance of modality-specific risk scores and integrated multi-modal stratification for incident ASCVD. (A) ROC curves for the four modality-specific risk scores in the held-out test set. (B) Uno’s time-dependent AUC across 10 years of follow-up. (C) Distributions of risk scores among ASCVD cases and controls. (D) Pairwise associations among modality-specific risk scores. (E) Independent hazard ratios from a multi-modal Cox model incorporating all four risk scores. (F) Change in model fit after removing each modality from the full Cox model. Excluding ProRS resulted in the largest reduction in log-likelihood, demonstrating its dominant incremental predictive value. (G) Absolute ASCVD event rates across ProRS × PRS tertiles. (H) Kaplan-Meier survival curves for joint ProRS-PRS risk groups.

Pairwise associations across the four modality-specific risk scores revealed distinct correlation structures, reflecting the heterogeneous biological and clinical information captured by each modality (Figure 2D). ProRS showed minimal correlation with the PRS (Spearman ρ = 0.04), indicating that proteomic signatures of ASCVD risk were largely independent of inherited genetic susceptibility. A similarly weak association was observed between PRS and the Clinical RS (ρ = 0.04), consistent with the interpretation that common genetic variation explains only a modest component of short- to mid-term ASCVD risk compared with contemporaneous clinical factors. In contrast, ProRS exhibited a strong positive association with the Clinical RS (ρ = 0.75), reflecting shared information between proteomic alterations and established cardiometabolic risk factors such as blood pressure, diabetes, and BMI. MetRS demonstrated a moderate-to-strong correlation with ProRS (ρ = 0.56), suggesting that proteomic and metabolomic pathways jointly capture overlapping but non-identical biological processes relevant to ASCVD development.

These correlation patterns demonstrate distinct and complementary roles of the four modalities in capturing ASCVD risk. The PRS represents a genetically anchored component of risk that is largely independent from present-day physiological status, as reflected by its minimal correlation with proteomic, metabolomic, and clinical measures. In contrast, proteomic signals show strong alignment with clinical risk factors, indicating that circulating proteins capture active cardiometabolic perturbations closely tied to blood pressure, metabolic health, and inflammatory processes. Metabolomic profiles occupy an intermediate position: they correlate moderately with proteomics, suggesting shared biological pathways, yet retain partial independence from clinical variables and almost complete independence from inherited genetic risk. Together, these relationships illustrate how each modality contributes a distinct dimension of risk—genetic predisposition, current physiological state, and metabolic pathway activity—underscoring the potential value of integrating multi-modal information for ASCVD prediction.

3.3. Combined Cox Modeling of Multi-Modality Risk Scores

To quantify the independent and combined prognostic contributions of each modality, the four standardized risk scores (PRS, Clinical RS, ProRS, and MetRS) were integrated in a Cox PH model fit in the training subset, and prognostic performance was evaluated in the held-out test set. All four modalities remained independently associated with incident ASCVD after mutual adjustment (Figure 2E). ProRS exhibited the strongest independent association (HR = 1.596, 95% CI: 1.424-1.789), followed by the PRS (HR = 1.307, 95% CI: 1.178-1.449) and the Clinical RS (HR = 1.237, 95% CI: 1.103-1.387). The MetRS had a weaker association (HR = 1.114, 95% CI: 0.990-1.254) and was not statistically significant at the conventional threshold.

Nested-model likelihood analyses further highlighted the dominant contribution of proteomics to ASCVD prediction (Figure 2F). Removing ProRS produced the largest reduction in model fit (Δlog-likelihood = 28.6), exceeding the impact of removing the PRS (12.4), Clinical RS (6.3), or MetRS (1.7). These results demonstrate that, among the four modalities, circulating proteins provide the most substantial incremental information for predicting future ASCVD events, even after accounting for genetics and clinical variables.

Discriminative performance analyses using the C-index corroborated these findings (Table 1). The 95% CI for the C-index were estimated by bootstrap resampling of the held-out test set with replacement, using percentile-based intervals. Using the PRS-only model as a reference (C-index = 0.576), the addition of Clinical RS or ProRS produced substantial improvements (C-index = 0.704 and 0.710, respectively). MetRS provided a modest gain (C-index = 0.651). Combining modalities yielded the strongest prediction: the model integrating PRS, Clinical RS, and ProRS achieved the highest C-index (0.722), marginally exceeding the full model that incorporated all four modalities (0.720). These trends indicate that proteomics provides the most powerful enhancement to genetic risk prediction, and that integration of genetic, clinical, and proteomic data provides the best overall discrimination.

Table 1.

Discriminative performance of PRS-only and multi-modality Cox models. C-index values are shown for models incorporating different combinations of modality-specific risk scores. The PRS-only model served as the baseline. Adding Clinical RS or ProRS substantially improved model discrimination, whereas MetRS provided a modest incremental benefit. The combination of PRS, Clinical RS, and ProRS yielded the highest C-index, slightly exceeding the performance of the full model.

These results show that genetics, proteomics, metabolomics, and clinical factors contribute complementary dimensions of ASCVD risk. The combined Cox framework demonstrates that proteomic signatures are the strongest independent predictors, while genetics captures orthogonal inherited susceptibility and clinical variables reflect contemporaneous physiological risk. Integrating these modalities yields a more comprehensive model of ASCVD risk than any single modality alone.

3.4. Multi-Modal Risk Stratification Using ProRS and PRS

To evaluate how complementary sources of molecular and genetic information jointly shape ASCVD risk, we examined stratification using ProRS and PRS—the two modalities that exhibited the most distinct and least correlated patterns in prior analyses. ProRS demonstrated the strongest independent association with incident ASCVD while PRS contributed additional but orthogonal prognostic information, indicating that proteomic dysregulation and inherited susceptibility represent biologically independent axes of risk. The minimal correlation between PRS and ProRS, combined with their additive effects in the Cox model, suggests that high-risk individuals may arise from elevated genetic risk, elevated proteomic signals reflecting active pathophysiology, or a combination of both. This motivated a focused analysis of joint ProRS-PRS stratification to determine whether combining these two modalities provides clearer separation of absolute ASCVD event rates and more refined identification of clinically meaningful subgroups compared with single-modality approaches.

Joint ProRS × PRS stratification revealed substantial heterogeneity in ASCVD event rates across the nine combined strata (Figure 2G). Event rates ranged from 4 percent in participants with both low ProRS and low PRS to 24 percent among those with high levels of both scores. Notably, individuals with high ProRS exhibited elevated risk regardless of genetic predisposition, with ASCVD event rates of 15%, 21%, and 24% across the low, medium, and high PRS tertiles, respectively. In contrast, high PRS alone did not confer equivalently elevated risk in the absence of proteomic abnormalities; participants with high PRS but low ProRS experienced an event rate of only 7%. These results show that proteomic dysregulation is a more proximal indicator of ASCVD risk, while genetic susceptibility modifies but does not fully determine clinical outcomes.

Survival analyses further illustrated how proteomic and genetic information contribute complementary insights into ASCVD risk (Figure 2H). Over 10 years of follow-up, the group with both high ProRS and high PRS showed the greatest decline in ASCVD-free survival, with cumulative event rates of 1.0%, 6.0%, and 14.8% at 1, 5, and 10 years. In comparison, individuals with low ProRS and low PRS had substantially lower event rates of 0.1%, 1.4%, and 3.2% at the same time points. Groups with discordant profiles, consisting of high ProRS with low PRS or low ProRS with high PRS, displayed intermediate survival patterns, indicating partial but not complete attenuation of risk. Overall, these results show that ProRS captures current physiological processes associated with ASCVD development, whereas PRS reflects inherited lifelong susceptibility. When combined, the two modalities provide a more refined and clinically informative stratification of ASCVD risk than either modality considered alone.

4. Discussion

This study demonstrates that integrating genetic, proteomic, metabolomic, and clinical information yields a substantially more comprehensive framework for predicting incident ASCVD than any single modality alone. The multi-modal design enabled the identification of distinct and complementary dimensions of ASCVD risk, including inherited susceptibility represented by the PRS, active biological processes reflected in proteomic and metabolomic signals, and traditional cardiometabolic risk factors encompassed by clinical variables. By systematically comparing and integrating these modalities, the study provides a detailed characterization of how each contributes to ASCVD development and clarifies their relative prognostic importance.

4.1. Biological and clinical insights from modality-specific risk information

Each modality captured a distinct aspect of ASCVD risk. Proteomic signatures demonstrated the strongest independent association with incident ASCVD, supporting their role in capturing ongoing biological perturbations such as inflammation, vascular remodeling, endothelial dysfunction, and metabolic dysregulation. The strong correlation between ProRS and clinical risk factors suggests that circulating proteins capture physiological processes that are closely related to, and partly overlapping with, established clinical risk profiles. Accordingly, the relationship between ProRS and Clinical RS is better interpreted as shared captured variance in proximal disease biology rather than strict independence. In contrast, the PRS showed near-zero correlation with proteomic, metabolomic, and clinical scores. This pattern supports the interpretation that inherited genetic susceptibility reflects a stable, lifelong baseline risk that is largely independent of contemporaneous biological activity. Although the PRS alone provided modest discriminative performance, its statistically independent contribution in the multi-modal Cox model underscores the relevance of genetic predisposition even after accounting for downstream molecular and clinical processes. Metabolomic features contributed intermediate predictive value and showed moderate correlation with proteomic profiles. This suggests that metabolites may represent biochemical intermediates linking genetic risk, proteomic activity, and clinical phenotypes. Their weaker independent association in the multi-modal Cox model may reflect the broader and less pathway-specific nature of NMR-based metabolomic measurements. Nonetheless, metabolomics added measurable incremental prognostic information and therefore represents a biologically meaningful intermediate layer between genetic liability and overt clinical risk factors.

4.2. Integrated multi-modal modeling and the complementary role of proteomics and genetics

The integrated Cox model revealed that proteomics provides the largest incremental improvement in ASCVD risk prediction, followed by genetics and clinical variables. Removal of ProRS resulted in the greatest reduction in model fit, demonstrating that circulating proteins contribute prognostic information that is not captured by genetics or traditional risk factors. Importantly, the joint analysis of ProRS and PRS highlighted the biologically independent yet clinically complementary nature of molecular and genetic risk. Individuals with elevated ProRS exhibited high ASCVD event rates regardless of genetic background, supporting the view that proteomic dysregulation reflects active pathophysiology that directly precedes clinical disease. In contrast, individuals with high PRS but low ProRS did not experience comparably high event rates, indicating that genetic liability alone may not translate into short-term clinical risk in the absence of active biological perturbation. The combination of high ProRS and high PRS identified the subgroup with the greatest risk over 10 years, indicating that inherited susceptibility amplifies the consequences of proteomic dysregulation. These findings demonstrate that integrating independent axes of risk can reveal clinically significant subgroups that would not be identifiable through single-modality analyses. The survival divergence across the nine ProRS × PRS strata further illustrates that combining modalities provides refined risk stratification with clearer clinical interpretability.

4.3. Implications, Limitations, and Future Directions

This study illustrates the potential value of integrating genetic, proteomic, metabolomic, and clinical information for incident ASCVD prediction. Proteomics provided the strongest independent signal, indicating active cardiometabolic and inflammatory processes, whereas genetics contributed a stable background risk and clinical variables captured established physiological determinants of ASCVD. Together, these modalities provided a more comprehensive framework for ASCVD risk prediction than any single modality alone, supporting the potential role of multi-modal molecular data in precision prevention.

Several limitations warrant consideration. First, because multi-modal profiling is available only for a subset of UKB participants, the analytic sample was necessarily restricted, reducing overall sample size and event counts compared with single-modality studies. This constraint may affect the stability of feature selection and downstream model performance. Second, the analytic cohort was predominantly composed of individuals of European ancestry, and ancestry-stratified analyses were not performed. As a result, the transportability of modality-specific and integrated risk scores to other populations remains uncertain. Differences in allele frequencies, linkage disequilibrium structure, baseline ASCVD risk, treatment patterns, and molecular biomarker distributions may affect model calibration and discrimination across populations. Third, while sex was included as a covariate and contributed strongly to the Clinical RS, we did not perform sex-stratified modeling. Given known sex differences in ASCVD pathophysiology and clinical presentation, sex-specific models may yield more accurate or biologically informative predictions.

Another limitation relates to the modeling framework. In this study, modality-specific risk scores were constructed using logistic regression to provide a consistent and interpretable summary of risk information within each modality, and these scores were then evaluated and integrated using survival-based analyses to account for censoring. This score-level integration strategy was intentionally chosen to improve interpretability and enable direct comparison of modality-level contributions within a unified framework. However, because the combined Cox model integrates modality-specific scores rather than raw features, it may not capture higher-order cross-modal interactions (e.g., protein-metabolite or genotype-protein interactions) that could further improve prediction. In addition, the current approach primarily supports modality-level association and contribution analysis rather than end-to-end multi-modal feature learning.

Additional data-related limitations should also be considered. Proteomic and metabolomic data were measured at a single baseline time point, preventing assessment of longitudinal molecular changes that may precede clinical ASCVD events. More frequent multi-omics sampling could reveal dynamic biological signatures that refine temporal prediction. Moreover, UKB participants are generally healthier than individuals enrolled in cardiovascular disease cohorts, which may reduce observed ASCVD incidence and attenuate effect-size estimates. Nonetheless, the strong performance of ProRS and the complementary contributions of multiple modalities within this relatively healthy cohort suggest meaningful translational potential for identifying early risk signals before overt disease onset.

Future work should evaluate this framework in ancestrally diverse and higher-risk populations, including ancestry-stratified and sex-stratified analyses, and assess whether recalibration, ancestry-specific modeling, or transfer-learning strategies are required for equitable implementation. Future studies should also compare score-based integration with survival-consistent score construction approaches (e.g., penalized Cox models for each modality) and unified feature-level multi-modal models that directly integrate raw genetic, proteomic, metabolomic, and clinical features. Such efforts may clarify condition-specific contributions of genetic risk, capture cross-modal biological interactions more explicitly, and further improve multi-modal ASCVD prediction.

5. Conclusion

In this study, we demonstrated that integrating genetic, proteomic, metabolomic, and clinical information provides a substantially more comprehensive assessment of ASCVD risk than relying on any single modality. Among the four modalities, proteomic profiles offered the strongest independent predictive signal, capturing active cardiometabolic and inflammatory processes that were not explained by inherited susceptibility or conventional risk factors. Genetic risk, while modest in magnitude, contributed orthogonal information and strengthened prediction when combined with molecular signatures. Metabolomic features provided intermediate but partially independent insight, whereas clinical variables captured established physiological determinants of ASCVD risk. The multi-modal Cox framework showed that each modality contributes a distinct component of ASCVD biology, such as lifelong inherited predisposition, current physiological state, and metabolic pathway activity, resulting in enhanced discrimination when integrated. Joint stratification by proteomic and genetic risk further revealed clinically meaningful subgroups, highlighting the potential value of combining molecular and genetic risk layers for precision prevention.

Overall, these findings suggest that large-scale molecular profiling, particularly proteomics, holds considerable promise for refining ASCVD risk prediction beyond traditional clinical and genetic models. Future efforts to incorporate longitudinal molecular measurements, evaluate diverse populations, and integrate modality-level signals directly at the feature level may further advance the accuracy and clinical utility of ASCVD risk stratification frameworks.