Systematic analyses uncover plasma proteins linked to incident cardiovascular diseases

Figures

• 

  Figure 1.. Graphical abstract. In our comprehensive study, we meticulously examined 2,920 plasma proteins measured longitudinally in a large cohort of 53,026 individuals from the baseline period of 2006–2010 to the censor date in September 2023, providing a robust median 14-year follow-up. Our investigation encompassed a spectrum of outcomes, including 14 incident CVDs, mortality, and various cardiac structural and functional parameters. Utilizing Cox models adjusted for demographic and lifestyle factors, we identified 892 proteins significantly associated with the 14 CVD outcomes and mortality after stringent Bonferroni correction. Additionally, we delved into the associations between the plasma proteome and 19 Cardiovascular Magnetic Resonance metrics using linear regression models. Employing a machine-learning-based prediction model trained on 257 pre-selected proteins and three sets of predictors (Protein, SCORE2, and Protein + SCORE2), we demonstrated the enhanced performance of plasma proteins in differentiating individuals developing CVDs from their healthy counterparts, thereby highlighting their supplementary contribution alongside established risk scales. Our investigation…

• 

  Figure 2.. Cox proportional hazard model correlates plasma proteome with CVD. (A) Numbers of significant associations between protein levels and 14 incident CVD outcomes. (B‒O) Volcano plots presenting the association of protein expressions with each outcome. The x-axis shows the HR, indicating the strength and direction of the association between protein levels and incident. The y-axis indicates the −log10 of the P-value for each association. Proteins with a P-value below the threshold determined by the Bonferroni correction (P < 0.05/2,920), are depicted above the black horizontal line. The red dots represent risk proteins, suggesting that they are significantly associated with an enhanced risk of incident CVD, while the blue dots represent protective proteins, indicating that they are significantly associated with a reduced risk of incident CVD. The top 8 proteins with highest −log10(P-value) for each outcome are marked with their names. The Cox proportional hazard models were adjusted for age, sex, ethnicity, TDI, blood collection season, time between protein measurement and blood sampling (in days), participant-reported fasting time, systolic blood pressure, BMI, smoking status, and…

• 

  Figure 3.. Linear regression links plasma proteins to cardiovascular magnetic resonance (CMR) metrics. (A) Manhattan plot displays the associations between plasma proteins and 19 CMR metrics. The significance threshold was adjusted to account for multiple comparisons, and unadjusted –log10(P-value) values are reported. Each circle represents a single protein-CMR association, with size indicating the regression coefficient The horizontal dashed line indicates that the Bonferroni-corrected threshold of –log10(P-value) is 4.76, above which the associations are considered significant. Dashed vertical lines separate five different phenotype categories: LV, RV, LA, and RA. (B) The heat map illustrates the association of each CMR metric with its most significantly correlated (top-15) plasma proteins. The color gradient from blue to red suggests the regression coefficient β. *, P < 0.05/2,920; **, P < 0.01/2,920; ***, P < 0.001/2,920. Multiple linear regression models were adjusted for age, sex, ethnicity, TDI, blood collection season, time between protein measurement and blood sampling (in days), participant-reported fasting time, systolic blood pressure, BMI, smoking status, and alcohol intake.

• 

  Figure 4.. Receiver Operating Characteristic analyses proved the feasibility of predicting incident CVD. (A) Bar plot displaying the AUCs of predictive models established based on three predictor sets: Protein (selected 257 plasma proteins), SCORE2 and Protein + SCORE2. Error bars represent the interquartile range derived by bootstrapping with 1000 iterations. ***, **, and * indicate significant difference tested using DeLong statistics with P-values < 0.001, < 0.01, and < 0.05, respectively. (B) Stacked bar chart showing protein importance calculated based on standardized information gain derived from LGBM classifiers. The top-15 proteins for each CVD outcome were selected and ranked based on their overall importance. Numbers on top of the bars indicate the amount of CVDs in which this protein ranked within the top-15 list. Thus we showed proteins that exhibited the most important discriminative value in two or more CVDs.

• 

  Figure 5.. Causal relationship of plasma proteins with incident CVDs and mediation effects of plasma proteins between modifiable factors and CVDs. (A) Manhattan plot showing the statistical significance (−log10[P-value]) of protein-cardiovascular disease associations derived from inverse variance weighted (IVW) Mendelian randomization analyses. Each point represents a unique protein-outcome association, with different colors indicating distinct cardiovascular outcomes. (B) The scatter plot illustrates the relationship between the number of associated cardiovascular outcomes (x-axis) and the median proportion mediated by proteins (y-axis). Each point represents a unique protein mediator, with point size and color intensity indicating the number of risk factors involved in the mediation pathway. (C) The bar plots indicate the five maximum amounts of mediation proportion for any proteins between risk factors and each CVD. The names of proteins and risk factors were annotated.

• 

  Figure 6.. The associations of CVD-associated proteins with genetic factors and functional enrichment analyses. (A) The associations of CVD-associated proteins with genetic risk for CVDs were assessed under five P-thresholds (pT_5e-06, pT_1e-05, pT_5e-05, pT_1e-04, pT_5e-04). Linear regression models, adjusted for age, sex, ethnicity, TDI, blood collection season, time between protein measurement and blood sampling, participant-reported fasting time, systolic blood pressure, BMI, smoking status, alcohol intake, and top ten genetic principal components, were utilized. Statistical significance was determined based on two-sided P-values, and to address multiple testing, Bonferroni correction was applied (P < 0.05/number of CVD-associated proteins). (B) The statistical significance level of each item is shown on the x-axis and is represented by −log10 of the q-value. y-axis shows different items, the bubble size indicates the enrichment count, and the bubble color indicates the different sources. (C) The protein importance is indicated by the color of nodes, with darker colors indicating higher SCORE2s. The width as well as gray value of the edges indicate the combined SCORE2 between adjacent…