Fig. 1

Flow chart for development and validation of DNAmCVDscore. Step 1: We train prediction models for developing DNAm surrogates for 13 CVD risk factors/biomarkers using data from the EPIC Italy study (n = 1803). We tested the validity of DNAm surrogates in four independent studies (n = 2107). Nine out of 13 DNAm biomarkers were validated in the testing set. Step 2: 60 candidate DNAm surrogates (nine newly developed + 51 from the literature) were regressed against the time from study recruitment to cardiovascular event in EPIC Italy (n = 1803). The elastic net regression model selected ten DNAm surrogates as components of the DNAmCVDscore. Step 3: In EXPOsOMICS CVD data set (N = 315), NICOLA (N = 1728), and HRS (N = 2146) we evaluated the prediction performance of DNAmCVDscore at different time points (right-censoring follow-up time) using logistic regression models adjusted for chronological age, sex, and recruitment centre (matching variables in EXPOsOMICS CVD) or Cox regression models (in NICOLA and HRS). DNAmCVDscore has a higher AUC for short-term cardiovascular events than for long-term CVD. Step 4: We compared the prediction performance of DNAmCVDscore with previously developed composite biomarkers: MRS, DNAmGrimAge, SCORE2 and SCORE2 + DNAmCVDscore. SCORE2 outperforms epigenetic predictors for long-term CVD risk (occurred more than 8 years after recruitment), whereas DNAmCVDscore predicts short-term events (occurred within 7 years after recruitment) better than other biomarkers. The enriched SCORE2 + DNAmCVDscore model outperformed all the competitors for the entire time horizon considered in the study