We conducted new DNAm assays of stored blood biospecimens collected from the DO-HEALTH randomized controlled trial (ClinicalTrials.gov identifier: NCT01745263) and merged these data with existing clinical data from the trial. Biospecimen assays were conducted blinded to the trial interventions and outcomes in the subgroup of 777 Swiss participants with samples available at baseline and year 3. The Cantonal Ethical Committee of the Canton of Zurich approved this study (BASEC-Nr 2021-02510). Written informed consent was obtained from all participants included in the study.
Study design and participants
This randomized, double-blind, placebo-controlled trial with a 2 × 2 × 2 factorial design had three primary treatment comparisons: (1) 2,000 IU per day of vitamin D compared to placebo; (2) 1 g per day of omega-3s (330 mg EPA plus 660 mg DHA from marine algae) compared to placebo; and (3) a strength-training exercise program performed for 30 min three times per week compared to an attention control exercise program focused on joint flexibility performed for 30 min three times a week. The factorial design was chosen to evaluate both the main and combined effects of the interventions. The inclusion criteria were age 70 years and older, living at home, having no major health events (no cancer or myocardial infarction) in the 5 years before enrollment, having sufficient mobility to visit the study centers without help and having good cognitive function with a Mini-Mental State Examination score of at least 24 (refs. 22,26). Participants did not receive any compensation for their involvement in the trial. The study and biobank are described at https://do-health.eu.
Randomization and masking
After enrollment and baseline testing, participants were randomized to one of eight treatment groups (Fig. 1) using block randomization (block sizes of 16 individuals) stratified by recruitment center, prior falls, sex and age (70–84 years or ≥85 years). A central randomization center in Switzerland, supported by trial software, was responsible for the blinding, treatment allocation and study intervention labeling. Participants received two gel capsules per day (vitamin D or placebo and omega-3s or placebo), identical in size, appearance, taste and weight. All capsules had coatings to prevent unblinding by aftertaste22,26.
Procedures
Participants were followed up for 3 years, both with yearly clinical visits (baseline and 1, 2 and 3 years) and telephone calls every 3 months. For DO-HEALTH Bio-Age, the subgroup of 777 Swiss DO-HEALTH participants was assessed at baseline and 3 years. The trial was performed at seven centers in five countries (Switzerland, France, Germany, Portugal and Austria). The study protocol and statistical analysis plan were approved by ethics and regulatory agencies in all five countries and have been previously published22,26. A data and safety monitoring board oversaw the study. Adherence to study medication was high, with 86% of the participants taking at least 80% of their total study pills; 70% of the participants performed the exercise programs at least twice per week, and 62% performed the exercise programs at least three times per week. High adherence to study medications (omega-3 and vitamin D supplements) was confirmed by changes in serum polyunsaturated fatty acid and 25(OH)D levels22,26.
DNAm data
Whole blood samples from study participants were collected in PAXgene DNA tubes and registered at the DO-HEALTH Biobank of the University of Zurich. Blood aliquots were sent to the Life&Brain Center, Department of Genomics, University of Bonn, on dry ice for DNA extraction with a chemagic magnetic beads-based method. DNA aliquots were processed on the Infinium Methylation EPIC version 1.0 array (Illumina). The EPIC array quantifies 5-methylcytosine levels at >850,000 CpG sites across all known genes, regions and key regulatory regions. Briefly, a 500 ng amount of extracted DNA samples was bisulfite converted using the EZ-96 DNAm lightning kit (Zymo Research), and 200 ng of the converted DNA was used as input for the EPIC arrays (Illumina). EPIC arrays were processed according to the manufacturer’s instructions and scanned using the Illumina iScan platform. The baseline and 36-month samples from the same individual were processed in the same array batch and on the same BeadChip to minimize batch effects. Quality control and normalization analyses were performed using the minfi (version 1.42.0) Bioconductor (version 2.46.0)55 package for the R statistical programming environment (version 3.6.3). Probes were considered missing in a sample if they had detection P values of >0.01 and were excluded from the analysis if they were missing in >50% of the samples. Normalization to eliminate systematic dye bias in two-channel probes was carried out using the minfi default method. Following quality control and normalization, DNAm data for 866,238 CpGs were available for 777 participants of the Swiss subgroup in DO-HEALTH, both at baseline and 3 years of follow-up (Fig. 1). Beta values were extracted and used for the analysis.
Epigenetic clocks and DNAm-based protein measures
Epigenetic clocks are DNAm algorithms that combine information from measurements across the genome to quantify variation in biological aging1. The first-generation epigenetic clocks were developed by comparing DNAm between individuals of different chronological ages to generate age prediction algorithms. We report analyses of two well-known first-generation clocks, Horvath’s multitissue clock2 and the blood-based clock developed by Hannum et al.3. These clocks are highly accurate in predicting chronological age but show only weak and inconsistent associations with morbidity and mortality2,3,7. The second-generation epigenetic clocks were developed by modeling differences in mortality risk and normalizing predicted risks to age values. We analyzed two second-generation clocks: PhenoAge4 and GrimAge2 (ref. 8). These second-generation clocks have much greater predictive capacity for morbidity and mortality than the first-generation clocks56.
The DunedinPACE pace-of-aging clock6 models differences in the rate of deterioration in organ system integrity, termed ‘pace of aging’ (ref. 57). In contrast to first- and second-generation DNAm clocks, which aim to quantify the amount of biological aging at the time of measurement, pace-of-aging clocks quantify the pace of age-related deterioration of system integrity. The DunedinPACE DNAm algorithm was derived from elastic net regression of the physiological pace-of-aging composite on Illumina EPIC array DNAm data derived from blood samples collected at the follow-up assessment at age 45 years. The CpG sites included in the DNAm dataset used to develop the DunedinPACE algorithm were restricted to those showing acceptable test–retest reliability as determined in the analysis by Sugden et al.58. The DunedinPACE DNAm algorithm is described in detail in the paper by Belsky et al.6.
For the Horvath, Hannum, PhenoAge and GrimAge clocks and the DNAm-based proteins underlying the GrimAge clock, we analyzed versions constructed from DNAm PCs, which have superior technical reliability compared to the original versions of these measures. This was achieved using the computational method by Higgins-Chen et al.29. The original versions of GrimAge2 and DunedinPACE demonstrate strong technical reliability6. Optimal test–retest reliability is a critical feature of measurements used to evaluate the impact of an intervention.
To compute GrimAge2, we submitted selected CpGs to the DNAm clock calculator hosted by the Horvath laboratory ( last accessed March 25, 2024) to derive the five DNAm clock estimates and DNAm protein estimates. After processing by the clock calculator (n = 790; Fig. 1), a complete sample of n = 777 with baseline and follow-up measurements was available for analysis.
The PC versions of the DNAm clocks (except GrimAge2) were calculated using the R code hosted on GitHub (https://github.com/MorganLevineLab/PC-Clocks)29 with R (version 4.2.1).
DunedinPACE was calculated for the same samples as the other DNAm variables according to the method described by Belsky et al.6, using the R code hosted on GitHub ( with R (version 4.2.1).
We first calculated the residuals of the regression of chronological age and biological age and then computed the difference between baseline and year 3. This difference was standardized (mean = 0, s.d. = 1) and was the outcome of our trial (Δyear 3−baseline).
Analysis
The analysis included all Swiss DO-HEALTH participants with available DNAm data at the trial baseline and the 3-year follow-up. We computed standardized change scores for all DNAm measures by comparing standardized residuals at the 3-year follow-up to those at baseline. We analyzed these change scores to test the hypothesis that omega-3 and/or vitamin D and/or exercise slow biological aging by conducting an intention-to-treat analysis comparing the change scores between participants randomized to the interventions and the control groups. We used analysis of covariance models for each of the DNAm measures. If there was no interaction between treatments, we quantified treatment effects as ‘main effects’ (for example, receiving vitamin D versus not receiving vitamin D while adjusting for other treatments). Alternatively, in the case of interactions between treatments, each treatment arm was compared to the placebo arm. The models were adjusted for chronological age (continuous and spline at 85 years), sex, history of falls before study enrollment (a stratifying variable of the trial), BMI and study site. The primary predictors were the three treatments and their interaction with time. Given multiple testing across four second-generation clocks, we focused on identifying consistent patterns in which changes from baseline to follow-up at year 3 had 95% CIs that did not include zero.
Standardization of effect sizes
Standardization (or normalization) is useful because it allows the comparison of epigenetic biomarkers measured on different scales by placing them on a common scale.
The SAS code below standardizes the change scores (differences) for the specified variables from baseline to the 3-year follow-up. By standardizing, each measure of epigenetic age acceleration (EAA) will have a mean of 0 and an s.d. of 1. We used the PROC STANDARD procedure in SAS to standardize several measures of difference in EAA (residuals of chronological versus biological age). The following SAS code was used for this standardization:
PROC STANDARD DATA = eaa_data MEAN = 0 STD = 1 OUT = df2;
VAR d_EAA d_EAApheno d_EAAhannum d_EAAgrim d_eaagrim2 d_PoA;
RUN;
In this code, the d_ prefix indicates the difference between the follow-up and baseline measures. This standardization process ensures that each difference measure has a mean of 0 and an s.d. of 1, facilitating easier comparison and analysis.
Explanation:
-
1.
DATA = eaa_data. This specifies the input dataset named eaa_data. This dataset contains the variables that will be standardized.
-
2.
MEAN = 0 STD = 1. These options indicate that the standardized variables should have a mean of 0 and an s.d. of 1.
-
3.
OUT = df2. This specifies the name of the output dataset, which, in this case, is df2. The standardized variables will be stored in this new dataset.
-
4.
VAR. This statement lists the variables to be standardized. The variables in this code are listed below.
-
d_EAA
-
d_EAApheno
-
d_EAAhannum
-
d_EAAgrim
-
d_eaagrim2
-
d_PoA
-
The d_ prefix stands for ‘difference’, indicating that these variables represent the difference (or change) from baseline to the 3-year follow-up. For example, d_EAA = EAA.followUp − EAA.baseline.
Baseline descriptives
We analyzed data from all Swiss DO-HEALTH participants for whom blood DNAm data were available at baseline and the 3-year follow-up (n = 777). The participants had a mean age of 75.5 years (s.d. = 4.5 years), and 60% were women (Table 1). Overall, 52% met the Nurses’ Health Study definition of healthy agers; the baseline average 25(OH)D level was 23.6 ng ml−1 (s.d. = 8.4 ng ml−1); and the baseline blood omega-3 (DHA and EPA) levels were, on average, 94.3 ng ml−1 (s.d. = 40.1 ng ml−1). The average baseline BMI was 25.7 kg m−2 (s.d. = 4.0 kg m−2), and 88% were physically active (29% were active one to three times per week, and 59% were active more than three times per week) based on the well-validated Nurses’ Health Study physical activity questionnaire59. Extended Data Table 1 shows the comparison of the Swiss subset to the total DO-HEALTH sample and the Swiss participants who had incomplete DNAm data and thereby were dropped from this analysis.
DNAm clocks
For the Horvath, Hannum and PhenoAge clocks, we analyzed versions constructed from DNAm PCs, which have superior technical reliability compared to the original versions of these measures29. The original versions of the GrimAge2 and DunedinPACE clocks already demonstrate high technical reliability5,6,8. All clocks except for DunedinPACE were regressed on chronological age, and residual values were calculated for analysis. DNAm clocks estimate biological age with regard to an organism’s biological state in comparison to a reference population age in which this state would be typical (Extended Data Table 5).
PhenoAge clock
The PhenoAge clock is another DNAm-based biomarker developed based on the analysis of nine blood chemistry markers, age and mortality data from the US National Health and Nutrition Examination Surveys (n = 9,926 participants aged 18 years and older; 23 years of mortality follow-up); DNAm and blood chemistry data from the InCHIANTI (Invecchiare in Chianti) study (n = 912 participants aged 21–100 years); and data from the US Health and Retirement Study (n = 3,593 participants aged 51–100 years)4.
GrimAge version 1 clock
The GrimAge clock was developed as a composite biomarker of seven DNAm surrogates of seven plasma proteins5, a DNAm-based estimator of smoking pack-years, age, and sex in the Framingham Heart Study Offspring and Gen3 cohorts (n = 2,751 participants aged 24–92 years)5. GrimAge relies on the fact that some (but not all) plasma protein levels can be estimated based on cytosine methylation levels. We included in our analyses the seven DNAm-based GrimAge proteins that relate to kidney function, mitochondrial function, blood clotting and inflammation (Extended Data Fig. 1).
GrimAge version 2 clock
The same unique 1,030 CpGs were used to construct version 2 of GrimAge based on individuals aged between 40 and 92 years who contributed 13,399 blood samples across nine study cohorts8. GrimAge2 outperforms GrimAge version 1 in predicting mortality across multiple racial/ethnic groups and in terms of associations with age-related conditions such as coronary heart disease8.
DunedinPACE
Pace-of-aging measures estimate the rate of biological aging, defined as the rate of decline in overall system integrity. Pace-of-aging values correspond to the years of biological aging experienced during a single calendar year. A value of 1 represents the typical pace of aging in a reference population; values greater than 1 indicate a faster pace of aging; values lower than 1 indicate a slower pace of aging. Based on the analysis of the pace of aging in the Dunedin Study (n = 817 participants examined at ages 26, 32, 38 and 45 years), the pace of aging was measured from within-person change over time in 19 blood chemistry and organ function test metrics of system integrity6. DNAm was measured at age 45 years.
Statistics and reproducibility
No formal sample size calculation was performed specifically for this study. The Swiss National Science Foundation funded DNAm assays of samples collected at baseline and 36 months from the Swiss subset of DO-HEALTH participants. The study flowchart is presented in Fig. 1. Of the 1,006 Swiss participants, 790 provided consent for these analyses. Samples with insufficient DNA extraction for accurate DNAm measures and samples that showed a mismatch of predicted sex based on DNAm measures and reported sex were excluded. This resulted in a sample size of n = 777 participants. A preliminary power calculation, based on the number of participants with both baseline and year 3 blood samples and consent, indicated that this sample size would provide 90% power for detecting the anticipated effects. No technical replicates were run for the DNAm analysis of the EPIC array.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
link