Blood cell composition reveals distinct biological interpretation of DNA methylation age and age acceleration

Abstract

Background

Epigenetic clocks are widely applied biomarkers of biological age, but their biological underpinnings remain unclear. We previously showed that epigenetic clocks are affected by naïve and memory T cell proportions, suggesting blood cell composition as a potential driver. However, analysis of cell composition is complicated by collinearity between cell fractions.

Methods

We first develop a principal component analysis (PCA) method that is robust to collinearity of blood cell counts and show that this approach provides biologically meaningful insights. Next, we quantify the contribution of cell composition to DNA methylation age and age acceleration estimated by six 1st - or 2nd -generation epigenetic clocks in an analysis of 4,058 samples. Finally, we test the influence of cell composition on the association between age acceleration and 176 incident health outcomes using a study of 18,859 individuals.

Results

We find that up to 53% of the variation in DNA methylation age can be attributed to cell counts, with a particularly strong contribution of the balance between naïve and memory T cells. Associations between age acceleration and cell counts are weaker (up to 21% variance explained) and, for 2nd -generation clocks, primarily involve neutrophils. We validate the contribution of cell counts to epigenetic clocks using artificial cell mixtures. Interestingly, cell composition significantly attenuates the association of age acceleration investigated with 176 incident health outcomes, although this effect remains modest.

Conclusions

DNAmAge and AgeAccel are markedly different in their associations with cell composition, indicating that they - at least in part - reflect different biological processes. Surprisingly, the variation in age acceleration that can be attributed to blood cell composition only has a minor contribution to its association with mortality and incident disease outcomes.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13073-026-01751-6.

Background

Aging is the leading risk factor for many diseases. However, due to the wide variation in health among older adults, researchers are increasingly interested in biomarkers that can differentiate biologically fit individuals from frail ones [1]. One class of aging biomarkers are epigenetic clocks, which use age-related changes in DNA methylation (DNAm) to estimate calendar age and/or age-related clinical biomarkers. Over the past decade, many clocks have been developed. First-generation clocks, such as the blood clock developed by Hannum et al. [2], the multi-tissue clock developed by Horvath [3], and the blood/saliva clock developed by Zhang et al. [4] predict calendar age. Second-generation clocks, such as PhenoAge [5], GrimAge [6], and DunedinPACE [7], predict composite biomarkers comprised of age and clinical measurements that are more directly indicative of health.

In aging research, epigenetic clocks are usually applied by regressing DNA-methylation age (DNAmAge) on calendar age to obtain age-acceleration residuals (AgeAccel). AgeAccel is commonly interpreted as an indicator of “biological age” [8]. Under this interpretation, a positive AgeAccel means that one’s “biological age” exceeds their chronological age compared to others in the cohort, thus the person is prematurely aging. Indeed, AgeAccels as calculated by epigenetic clocks have been linked to numerous age-related health outcomes, including all-cause mortality, cardiovascular disease, cancer, and diabetes [9]. Various commercial parties now sell kits to measure DNAmAge, claiming, for instance, that it is “the age that matters” [10] or one’s “real age” [11].

Despite the widespread application of epigenetic clocks, the concept of “biological age” remains ill-defined, and the biological mechanisms driving these clocks are largely unknown. We previously reported that first-generation clocks detect age-related changes in the proportions of naïve and activated lymphocytes in blood, with predicted ages varying by up to 40 years between purified naïve and activated T cells and NK cells from the same donor [12]. However, a specific and quantitative analysis of the contribution of cell proportions to DNAmAge and AgeAccel from 1st and 2nd generation clocks is currently lacking.

Recent advances in cell deconvolution algorithms, such as EpiDISH [13] and IDOL-extended [14], enable accurate estimation of high-resolution cell proportions in blood samples from DNAm-data, which can subsequently be tested for associations with an outcome of interest [13]. However, the analysis of cell counts is challenging due to collinearity between individual fractions, meaning that cell proportions cannot change independently. In addition to the biological correlations between fractions, they are also mathematically constrained to sum to 100%. As one cell fraction increases, another must decrease to maintain this total, making it impossible to estimate the effects of changes in individual cell fractions and health outcomes. Novel approaches are needed to address the issue of collinearity.

In this study, we first outline the challenges collinearity poses when analyzing cell fractions and introduce a principal component analysis (PCA) approach to address this. We then apply this method to a dataset of 4058 individuals to investigate the associations of cell fractions with six epigenetic clocks, including 1st and 2nd generation clocks, both DNAmAge and AgeAccel. We find striking differences between DNAmAge and AgeAccel in both the strength of their associations with cell composition and the specific cell fractions involved. We validate these results in an external dataset of in silico generated artificial cell mixtures. Finally, we use data on 18,859 individuals to show that adjustment for cell composition PCs modestly attenuates the association of AgeAccels with 176 incident health outcomes. We conclude that the biology underlying variation in DNAmAge and AgeAccel is distinct from the perspective of cell composition.

Methods

Cohorts

This study was performed using DNAm data generated within the Biobank-based Integrative Omics Studies Consortium (BIOS Consortium). All data were generated in whole blood samples originating from 6 Dutch biobanks: Cohort on Diabetes and Atherosclerosis Maastricht (CODAM) [15], LifeLines (LL) [16], Leiden Longevity Study (LLS) [17], Netherlands Twin Register (NTR) [18], Rotterdam Study (RS) [19], and the Prospective ALS Study Netherlands (PAN) [20]. NTR is a biobank of twins, so a random person was selected from each twin pair to ensure that samples were unrelated. We selected samples from all cohorts for which DNAm data passed quality control (QC), which was the case for 4058 samples. Calendar age of samples ranged from 18 to 87 (Fig. 1A, Additional File 1: Table S1). This dataset was used for all analyses with the exception of the association between epigenetic clocks and cell composition PCs and health outcomes.

Fig. 1.

Distribution of age and cell fractions. (A) Histogram of ages of the 4058 samples included in our study. (B) Pie chart of cell fractions in all 4058 samples. The percentages of each fraction are shown on the slices of the chart. Percentages must always add up to 100% for any given person. (C) Scatter plots of percentage per cell fraction against age. Pearson correlations (R) are shown for each fraction. Trendlines represent a linear model with the fraction as the outcome and calendar age as the determinant. Asterisks indicate Bonferroni-adjusted significance of age correlation (*: p_Bonf< 0.01, **: p_Bonf< 10^− 5, ***: p_Bonf< 10^− 10)

To test the association of epigenetic clocks and cell composition PCs with health outcomes, we used an external dataset including 18,859 participants of the Generation Scotland cohort. This cohort is described in more detail in a previous publication [21]. Briefly, this cohort measured the incidence of 176 health outcomes, including all-cause mortality, as well as DNAm using the Illumina EPIC array.

DNA methylation data

DNAm data were generated by bisulphite-converting 500 ng of genomic DNA from whole blood samples using the EZ DNA Methylation kit (Zymo Research, Irvine, CA, USA). Subsequently, bisulphite-converted DNA was hybridized onto Infinium HumanMethylation450 BeadChip arrays (Illumina, San Diego, CA, USA), and signal intensities were measured on an Illumina iScan BeadChip scanner according to the manufacturer’s protocol.

A detailed description and R code of the pre-processing workflow, including the quality control and normalization, is described in DNAmArray, which is publicly available at https://molepi.github.io/DNAmArray_workflow/. In brief, IDAT files were read into R using minfi [22], after which sample-level QC was performed with MethylAid [23]. Low-quality samples were defined based on 4 QC-measures using information provided by the control probes on the array and 1 measure was based on call rate (> 95%) (Additional File 1: Fig. S1), which led to the exclusion of 168 samples. Probe-level QC was based on detection p value (p < 0.01), number of beads available (≤ 2) or zero values for signal intensity. Resulting probes with more than 5% missing values were removed. 4121 samples passed QC. Normalization was done using functional normalization as implemented in minfi, using five principal components extracted using the control probes for normalization.

Batch effect correction

We observed batch effects in the DNAm-data between the 6 included biobanks, therefore we performed a correction for batch effects using the ComBat R package [24]. Age was taken as the variable of interest to prevent age differences between biobanks from being removed. To check if batch correction went correctly, PCA-plots were made of the DNAm-data before and after ComBat (Additional File 1: Fig. S2A). After ComBat, batch effects were resolved (Additional File 1: Fig. S2B). To validate this step, we repeated the analysis of Fig. 3 (effect of cell composition PCs on age) in each of the 6 biobanks separately, before and after applying ComBat (Additional File 1: Fig. S2C). The results look very similar for all cohorts, both before and after ComBat. For example, PC2 and PC6 have the strongest negative age effects in each cohort, while PC7 always has the strongest positive age effect. These results match the full analysis (Fig. 3A). We conclude that ComBat did not introduce spurious results.

Fig. 3.

Association of PCs calculated from cell fractions with age. (A) Scree plot of the PCs calculated from cell fractions. The variance explained of each PC is shown on each bar. (B) Heatmap of correlations between the 11 PCs and the 12 cell fractions from which they were calculated. (C) Effects of the 11 PCs on age. PCs were manually labelled based on the correlations observed in 3B. Effects are interpreted as: “If the PC goes up by 1 standard deviation, age goes up by X years.” (D) Effects of PCs on age when leaving PC2 out of the model. Effect sizes remain identical, with only marginal changes in the confidence intervals and a lower total explained variance

Cell composition

To estimate cell composition, we used the EpiDISH deconvolution algorithm as implemented in the EpiDISH R package [25]. We used the reference panel of 12 blood cell types: neutrophils, eosinophils, basophils, monocytes, naïve B cells, memory B cells, naïve CD4 T cells, memory CD4 T cells, and naïve CD8 T cells, memory CD8 T cells, regulatory T cells (Tregs), and natural killer cells (NK cells). For deconvolution, we used the RPC-method.

Outliers in cell composition were removed as follows: for each fraction, any sample which fell at least 5 SDs away from the mean was removed. Across all samples, this led to the exclusion of 63 samples, leaving 4058 samples for analysis.

Epigenetic clocks

Four 1st -generation epigenetic clocks were studied: the blood clock developed by Hannum et al. (Hannum) [2], the multi-tissue clock developed by Horvath (Horvath) [3], and the blood/saliva clock developed by Zhang et al. (Zhang) [4]. Additionally, we included three 2nd -generation clocks, all trained on blood samples: PhenoAge, which predicts a composite score of blood health biomarkers [5], GrimAge v2, which predicts a composite score of smoking and several plasma proteins [6], and DunedinPACE, which predicts a composite score of longitudinal changes in functional biomarkers [7]. All clocks are described in Table 1.

Table 1.

Epigenetic clocks included in our study

Epigenetic clock	Generation	Number of CpGs	Training samples	Training tissues	Publication date
Hannum	1st	71	656	Blood	2013 [2]
Horvath	1st	353	3931	27 tissue types	2013 [3]
Zhang	1st	514	13,566	Blood, saliva	2019 [4]
PhenoAge	2nd	513	456	Blood	2018 [5]
GrimAge	2nd	1030	1833	Blood	2022 [6]
DunedinPACE	2nd	173	1037	Blood	2022 [7]

Hannum, Horvath, and Zhang are first-generation clocks, trained to predict calendar age. PhenoAge, GrimAge, and DunedinPACE are second-generation clocks, trained to predict age-related composite biomarkers. GrimAge V2 was used

Hannum, Horvath, Zhang, and PhenoAge were calculated using the methylclock R package [26]. For GrimAge, the methylCIPHER package was used [27]. For DunedinPACE, we used the DunedinPACE R package (https://github.com/danbelsky/DunedinPACE).

We calculated age acceleration scores (AgeAccel) from epigenetic clocks by taking the residuals of a linear model taking calendar age as the determinant and DNAmAge as the outcome:

Notably, DunedinPACE is unique in the fact that instead of estimating DNAmAge, it estimates the pace of aging, i.e. the rate at which physiological function is deteriorating for a person. For this reason, we transformed its score, which is in years by default, into weeks, to make its scale match the other clocks more closely.

Age and cell counts

A series of example analyses were performed to illustrate how collinearity between cell fractions affects models attempting to use cell fractions as predictors for an outcome, either in univariable or multivariable approaches. First, 12 univariable models were run, using age as the outcome and a single fraction as the predictor:

To illustrate that collinearity between cell fractions causes their effect sizes to be intertwined, we repeated this model 11 times for the neutrophil fraction, each time adjusted for one of the 11 other fractions:

Subsequently, to demonstrate the effects of collinearity on multivariable models, 12 variations of the same multivariable models were run, using age as the outcome and all 12 fractions as predictors:

This model was repeated 12 times, each time coding a different fraction as the 12th. Since the effect size of 12th fraction cannot be estimated due to collinearity (see Results), the 12th fraction is automatically excluded from the model and incorporated into the intercept.

To solve the issue of collinearity, we calculated principal components from the cell counts. Since the cell composition contains 11 independent variables (the 12th fraction being identical to 100% minus fractions 1–11), the PCA resulted in 11 PCs. PCs were labelled based on their correlations with individual cell counts to give them biological meaning (Fig. 3B). To check whether PC1 was related to smoking, an ANOVA was performed in the 3181 samples whom smoking status was known, using smoking as the predictor and PC1 as the outcome (Additional File 1: Fig. S3), indicating no association. We then ran a model using these PCs as age predictors:

Epigenetic clocks and cell counts

We analyzed the association between cell counts and DNAmAge/AgeAccel according to the six investigated clocks by running a linear model using the cell count PCs as predictors and DNAmAge or AgeAccel of each clock as the outcome:

To find the variance explained by each individual PC, we ran a series of univariable models and obtained the R²-value:

Notably, because PCs are orthogonal to each other, the effects obtained from a univariable model are identical to those from a multivariable model. To test whether these PC effects were influenced by smoking, we regressed the PCs on smoking status in the 3181 samples for whom smoking status was known and repeated the DNAmAge/AgeAccel analyses (Additional File 1: Fig. S4).

As an alternative solution to collinearity, we tested the effect of raising one fraction by 1 SD while proportionally shrinking the others on AgeAccel as calculated by the clocks. To do this, we first ran a multivariable model without an intercept to model the association with AgeAccel (for each clock separately) for all cell types:

Then, we predicted AgeAccel from the average cell composition observed in Fig. 1B. Additionally, we predicted AgeAccel 12 more times, each time adding 1 SD to one fraction while proportionally shrinking all others to maintain the 100% total. To obtain the effect of the cell count perturbations on AgeAccel, we subtracted the AgeAccel from the average cell composition (Fig. 1B) from the AgeAccel obtained from the 12 modified cell compositions (AgeAccel_x - AgeAccel_avg).

Finally, to validate our results using an external dataset, we generated artificial cell mixtures in silico. DNAm-data from sorted cell types were downloaded from GEO (GSE167998). In total, this resource contained DNAm-data for 56 sorted blood samples. The number of samples per cell type varied from 3 to 6, and samples from different cell types originated from different donors with different ages (age range 19–58). Per cell type, we calculated the average age/DNAm profile (avgCT) by taking the mean age and beta-value per CpG. Subsequently, we constructed artificial cell mixtures in silico by taking the population mean of cell type composition as found in Fig. 1B (e.g. 56.8% Neu), combining the age and DNAm-profiles of each avgCT in a weighted average. Then, we made 12 more artificial samples, each time raising the percentage of 1 avgCT by 1 SD while proportionally shrinking the others to maintain the 100% total. At this point, we were left with 13 artificial samples, each with a mean age and a certain DNAm-profile. From these DNAm-profiles, we calculated the clocks’ DNAmAges. AgeAccel measures were obtained by subtracting the mean age of the artificial sample from its DNAmAge.

Association with health outcomes

We adapted the analysis published by Mavrommatis et al. [21] to the association of epigenetic clocks and cell composition PCs with 176 incident health outcomes in 18,859 participants included in the Generation Scotland cohort. The number of disease outcomes studied differs from the previous study due to the inclusion of Scottish Cancer Registry (SMR06) records in addition to the general/acute and inpatient day case records (SMR01). To test the effect of cell composition as a whole on health outcomes, we ran two Cox regression models for each outcome: a null model only including known demographics (age, sex, BMI, smoking pack years, alcohol intake, years of education, and socioeconomic deprivation score) as determinants, and another model also containing all 11 cell composition PCs. The goodness of fit of the two models were compared using a likelihood ratio test (LRT). To test the effect of each individual PC on each outcome, a Cox regression model was run including the PC and demographics as determinants. Finally, to test the effect of adjusting for cell composition PCs on the association of AgeAccel on health outcomes, we ran two Cox regression models for each clock-outcome pair: one including the clock and demographics as determinants, and one also including all 11 cell composition PCs. Hazard ratios (HRs) of AgeAccel were extracted from both models and compared.

Results

Dataset and cell counts

Our dataset comprised 6 Dutch cohorts participating in the Dutch BIOS Consortium (Additional File 1: Table S1) for which DNA-methylation (DNAm) was measured using the Illumina 450 K-array. We analyzed 4058 whole blood samples across an age range of 18–87 (Fig. 1A). We used EpiDISH to estimate the proportions of 12 blood cell types (Fig. 1B): neutrophils, eosinophils, basophils, monocytes, naïve and memory subsets of B-cells, CD4 T cells, and CD8 T cells, regulatory T cells (Tregs), and natural killer cells (NK cells). Neutrophils were the largest fraction, making up 56.8% of the total blood composition on average (Fig. 1B). Exploratory analysis revealed age correlations for most fractions (Fig. 1C). For example, naïve T cells showed negative correlations with age (R_{naïve CD4 T cells} = -0.24, R_{naïve CD8 T cells} = -0.59), whereas memory T cells showed positive age correlations (R_{memory CD4 T cells} = 0.23, R_{memory CD8 T cells} = 0.17).

Analysis of cell fractions is complicated by collinearity

Before considering epigenetic clocks, we explored the effects of collinearity by investigating the link between cell fractions and calendar age using a univariable regression model for each fraction. Significant associations with age (P_Bonferroni < 0.01) were found for all fractions except neutrophils and eosinophils (Fig. 2A). Strong positive associations were observed for basophils, memory B cells, and memory CD4 T cells (effect size = 1.4, 1.4, and 1.2 years/%, respectively), while naïve CD8 T cells, regulatory T cells, and naïve CD4 T cells showed the strongest negative associations (effect size = -3.9, -2.8, and − 1.1 years/%, respectively).

Fig. 2.

Collinearity complicates univariable or multivariable analysis of individual cell fractions. (A) Effects of 12 blood cell fractions on age from the univariable analysis. Each fraction’s effect on age was estimated independently, without considering other fractions. Effect sizes are interpreted as: “for every percentage point increase in the fraction, age increases by X years.” (B) Pearson correlations between the cell fractions. (C) and (D) Effects of the 12 blood cell fractions on age from the multivariable analysis including all fractions as predictors. Due to collinearity, one fraction must be left out, skewing the effect sizes of the remaining fractions. Panel C shows results excluding naïve CD8 T cells, and panel D shows results excluding basophils. The intercepts and the percentage of age variance explained by each model are shown at the top of each panel

However, cell fractions are correlated, both due to biological associations and because they mathematically sum to 100% (Fig. 2B). Neutrophils were weakly to moderately negatively correlated to all other fractions (R ranging from − 0.2 to -0.6), while the remaining fractions generally showed weak positive correlations (R up to 0.3). To assess how these correlations impacted the interpretation of effect sizes, we repeated the regression of the neutrophil fraction on age 11 times, adjusting for one of the 11 other fractions each time (Additional File 1: Fig. S5). The effect of the neutrophil fraction on age ranged from 0.05 (no adjustment) to 0.4 (adjusting for memory CD4 T cells) to -0.3 (adjusting for naïve CD8 T cells), illustrating how collinearity makes effect sizes ambiguous.

To account for correlations between fractions, we performed multivariable analysis by including all 12 fractions in a single model (Fig. 2C-D). However, since the 12 fractions add up to 100%, the 12th fraction contributes no independent information (it is equal to 100% minus fractions 1 through 11), thus its effect on age cannot be estimated (shown as dashed red lines in Fig. 2C-D). This 12th fraction is incorporated into the model’s intercept, skewing the effect sizes for the remaining fractions. For instance, if naïve CD8 T cells are integrated into the intercept, all other fractions show strong positive age associations (P_Bonferroni < 10^− 10 for all fractions, Fig. 2C). Conversely, if basophils are incorporated into the intercept, all other fractions show negative age associations (P_Bonferroni < 0.001 for all fractions, Fig. 2D).

We concluded that multivariable analysis cannot produce clear effect sizes for individual fractions due to collinearity. It should be noted that this issue does not affect the total effect of all 12 fractions combined, as all models explained the same proportion of age variance (43.5%, Fig. 2C-D and Additional File 1: Fig. S6). Therefore, the multivariable approach is valid for estimating the combined effect of cell fractions, such as when adjusting an analysis for cell composition.

PCA-based method enables collinearity-free investigation of cell composition

Principal component (PC) regression is a known approach for handling collinearity [28]. PCs are linear combinations of variables - in this context, cell fractions - that are constructed to be uncorrelated. Consequently, each PC can vary independently of the others.

We calculated 11 PCs from the cell composition data (Fig. 3A), corresponding to the 11 independent cell fractions. To interpret the biological meaning of these PCs, we correlated them to the individual cell fractions (Fig. 3B).

The top PCs could reliably be interpreted as either single fractions or balances between fractions. For instance, PC1 was very strongly correlated to neutrophils (R = 1.0), and negatively to most other fractions, so we labelled it ‘Neutrophils’. This shows that the primary axis of variation in cell counts is the proportion of neutrophils relative to all other fractions. PC2 was strongly positively correlated to naïve CD4 and CD8 T cells (R = 0.7 and 0.6, respectively) and strongly negatively correlated to memory CD8 T cells (R = -0.6), thus we labelled it ‘T cell naïve/memory’. This means that the second-strongest feature in blood cell counts is the balance between naïve and memory T cell phenotypes. We labeled PCs 1–6 and PC8 in this way. For PC7 and PCs 9–11, the correlations of PCs became diffuse (no absolute correlation above 0.5); these PCs were labeled as ‘Mixed’, with an addition of their strongest correlation.

Next, we used the 11 labeled PCs in a multivariable regression model to predict age (Fig. 3C). Changes in cell composition explained 43.5% of age variation (equal to the previous multivariate analysis), with the strongest associations observed for PCs representing naïve and memory T cells (PC2, PC6, and PC7, P_Bonferroni < 10^− 10). Specifically, PCs correlating positively with naïve CD8 T cells (PC2 and PC6) showed negative age association (effect size = -6.5 and − 6.3 years/SD, respectively), while the PC correlating negatively with naïve CD8 T cells (PC7) showed positive age association (effect size = 4.4 years/SD). This aligns with the known shift from naïve to memory T cells during aging [29]. Other PCs, such as those representing memory B cells vs. monocytes (PC4) and naïve B cells vs. eosinophils (PC5), also showed significant but smaller associations with age (effect size = 1.0 and − 2.0 years/SD, respectively). Interestingly, while the ratio between neutrophils and all other fractions (PC1) explained a large proportion of the variance in cell composition (33%, Fig. 3A), this PC was not associated with calendar age (Fig. 3C), indicating that variation in the neutrophil fraction is age-independent.

To demonstrate that the PC-approach is robust against collinearity, we repeated the age regression without PC2, which had the strongest age association (Fig. 3D). This pruned model’s intercept and PC effect sizes remained identical to the full model, and standard errors increased only marginally, and the total variance explained dropped to 17.4% (from 43.5%) due to the exclusion of an age-related PC. Taken together, the PC-approach provides statistically valid and biologically meaningful results when investigating the associations of cell fractions.

Naïve and memory T cells are strongly associated with age and DNAmAge but not AgeAccel

Next, we applied the PC-approach to study the association between cell fractions and several epigenetic clocks, considering both DNA methylation age (DNAmAge) and age acceleration residuals obtained by regressing DNAmAge on calendar age (AgeAccel). We included six clocks in our analyses, including three 1st -generation clocks (Hannum, Horvath, and Zhang) and three 2nd -generation clocks (PhenoAge, GrimAge, DunedinPACE) (Table 1). All clocks except DunedinPACE correlated very strongly with calendar age (R ≥ 0.94, Additional File 1: Fig. S7). This was expected, since these clocks are trained to predict age (1st -generation clocks) or composite biomarkers composed of age and clinical measurements (2nd -generation clocks). DunedinPACE, which predicts the rate of longitudinal change in age-related clinical measurements rather than biological age at a single timepoint, was correlated moderately with age (R = 0.43). AgeAccel was calculated for all clocks by linearly regressing DNAmAge on calendar age and obtaining the residuals from this model.

First, we considered Horvath’s multi-tissue clock (Fig. 4A). Since this clock’s DNAmAge correlates extremely strongly to calendar age (Additional File 1: Fig. S7), it showed very similar associations with cell fractions. The strongest correlations were once again found for PC2, PC6, and PC7 (effect size = -6.4, -5.5, and 3.4 units/SD, respectively), which represent naïve and/or memory T cell fractions. Notably, AgeAccel associations were much weaker for these PCs (effect sizes of -0.7, 0.1, and − 0.5 units/SD, respectively); PC6’s association became non-significant, and PC7’s association even reversed in direction.

Fig. 4.

Association of cell fraction PCs with epigenetic clocks. (A-F) Effects of the 11 cell fraction PCs on the DNAmAge (blue) and AgeAccel (red) of three first-generation clocks - Horvath, Hannum, and Zhang (A-C) and three 2nd-generation clocks - PhenoAge, GrimAge, Dunedin PACE (D-F). PCs were manually labelled based on the correlations observed in Fig. 2B. Effect sizes are interpreted as: If the PC goes up by 1 standard deviation, DNAmAge/AgeAccel goes up by X units (weeks for DunedinPACE, years for all other clocks)

The six clocks included in our study have different training tissues (e.g. blood samples for Hannum and 27 different tissue types for Horvath) and predict different outcomes (calendar age for 1st -generation clocks and health-related composite biomarkers for 2nd -generation clocks). As such, the clocks could capture different biological processes and thus associate differently with cell fractions. These differences, if any, would be mostly be observed in the clock’s AgeAccels, since the DNAmAges of all clocks except DunedinPACE are strongly correlated to calendar age and thus also to each other.

When examining AgeAccel across all clocks (Fig. 4), we observed both shared and distinct patterns. For all clocks, AgeAccel had weaker associations with cell composition than DNAmAge, especially for the PCs representing T cell subsets (PC2, PC6, and PC7). For instance, the effect size of PC2, representing T cell naïve-to-memory ratio, ranged from − 1.5 to -7.3 units/SD for DNAmAge, and from − 0.3 to -1.1 units/SD for AgeAccel, with Hannum and PhenoAge showing the greatest effects (effect size = -1.1 and − 0.9 units/SD, respectively). Interestingly, the effect of PC1, representing the neutrophil fraction, varied greatly between clocks, being minimal for Horvath and Zhang (effect size = 0.1 units/SD for both) but higher for Hannum, PhenoAge, GrimAge, and DunedinPACE (effect size = 0.6, 1.0, 0.7, and 0.9 units/SD, respectively).

In summary, we observed that the AgeAccels of Horvath and Zhang were mostly independent of cell fractions, with only minor contribution from T cell phenotypes, while GrimAge and DunedinPACE were mostly affected by the neutrophil fraction, and finally Hannum, PhenoAge were affected by both T cells and neutrophils. We repeated the analysis after regressing PCs on smoking status and found virtually identical effects (Additional File 1: Fig. S4A), indicating that smoking did not influence our results.

Effect sizes shown in Fig. 4 are dependent on the scale of each outcome. Therefore, we assessed the relative contribution of cell composition to the epigenetic clocks by estimating the proportion of variance of DNAmAge and AgeAccel that was explained by cell composition. Cell composition explained 44% of the variance in calendar age (Fig. 5A, panel “Age”) and likewise 44–53% of the variance for all DNAmAges except DunedinPACE, for which 26% was explained (Fig. 5A, panel “DNAmAge”). Again, the primary features explaining this variance (Fig. 5B, panels “Age” and “DNAmAge”) were PC2, PC6, and PC7, which all represent naïve and memory T cells. Other PCs explained only marginal proportions of the variance (≤ 2.3%), except for DunedinPACE, for which 7.8% of variance was explained by PC1 (representing neutrophils).

Fig. 5.

The proportion of variance of epigenetic clocks that can be explained by cell composition. (A) Variance explained by cell composition for age (black bar) and DNAmAge and AgeAccel of each clock investigated (blue and red bars respectively). Numbers show the total percentage of the given outcome that can be explained by cell composition. (B) Variance explained by each individual cell composition PC. Numbers in each cell show the percentage of the given readout (columns) that can be explained by one cell composition PC (rows)

The relationship between cell composition and AgeAccel varied greatly across clocks (Fig. 5A, right panel). For example, cell composition explained just 4% of the variance in Zhang’s AgeAccel, but 21% in Hannum’s AgeAccel. This suggests that while DNAmAge behaves similarly across clocks (except DunedinPACE), AgeAccel has different biological interpretations, with some being more affected by cell composition than others.

Not only the total of the variance explained by cell composition differed for the clocks’ AgeAccels, but also the contributing cell fractions (Fig. 5B, panel “AgeAccel”). The AgeAccel of Hannum and PhenoAge associated primarily with PCs representing neutrophils (PC1) and T cells (PC2, PC6, and PC7), while the AgeAccel of GrimAge and DunedinPACE associated mostly with neutrophils (PC1), with minor contribution from other PCs (PC2, PC5, PC6, PC7, PC8). In line with previous analyses, cell composition barely explained the AgeAccels of Horvath and Zhang, with only minor contributions from PC2 (T cell naïve-to-memory ratio).

The divergent associations of calendar age and AgeAccels suggest that CpGs comprising the various clocks could be consistent calendar age trackers, while others may be more informative for deviations from the norm for a given age. To explore this, we analyzed how a 1-year increase in either calendar age or AgeAccel affects the DNAm values of the CpGs making up the clocks (Additional File 1: Fig. S8). We observed that the CpGs comprising first-generation clocks (Hannum, Horvath, Zhang, Additional File 1: Fig. S8A-C) showed strong similarity in their effect sizes for calendar age and AgeAccel (R = 0.91, 0.80, and 0.88, respectively), while CpGs comprising second-generation clocks (PhenoAge, GrimAge, DunedinPACE, Additional File 1: Fig. S8D-F) had more distinct effects on the two (R = 0.50, 0.37, and 0.13 respectively). Strikingly, the effect estimates for AgeAccel were much less precise than those for calendar age, with standard errors being 2.8–3.5 times higher (Additional File 1: Fig. S8). These results suggest that AgeAccels contain considerable noise, and the AgeAccels of 2nd -generation clocks in addition capture information that is distinct from calendar age.

Taken together, the association of DNAmAge and AgeAccel with cell fractions are strikingly different, both in magnitude and in terms of the cell fractions that contribute. Secondly, the AgeAccels of the various clocks differ in their association with cell fractions, suggesting that different clocks have different biological underpinnings. These associations remained virtually identical after adjusting for smoking (Additional File 1: Fig. S4B).

Artificial cell mixtures suggest that cell composition drives epigenetic clocks

Another approach to address collinearity is to model the effect of an increase in one fraction while proportionally reducing all other fractions. This method ignores biological correlations between fractions, but does address the fractions summing to 100%. To apply this method, we computed the average cell composition for our study (Fig. 1B), and predicted the clocks’ DNAmAges for this profile as a reference. Next, we systematically created artificial cell profiles by raising one fraction by 1 standard deviation, while shrinking all other fractions proportionally to maintain the total at 100%. We then recalculated DNAmAges to analyze the AgeAccel caused by these shifts in cell composition (Fig. 6A).

Fig. 6.

Effect of cell composition changes on AgeAccel. (A) The change in AgeAccel observed after adding 1 SD of a cell type to a baseline sample with mean cell count distribution while proportionally shrinking all others, using our dataset. The X-axis shows which cell type was raised for this sample. The Y-axis shows the AgeAccel for the six clocks compared to the baseline sample (not shown, set to 0 as a reference). AgeAccels are in units of years for all clocks except DunedinPACE, for which it is in weeks. (B) Generation of the artificial cell mixtures. DNAm-data from sorted cell types were downloaded from GEO (GSE167998, 56 samples total). Samples from different cell types originated from different donors with different ages (age range 19–58). Per cell type, we calculated the average age/DNAm profile (avgCT). Subsequently, we constructed artificial cell mixtures in silico by taking the population mean of cell type composition as found in Fig. 1B (e.g. 56.8% Neu), combining the age and DNAm-profiles of each avgCT in a weighted average. We then generated 12 more artificial samples, each time raising the percentage of 1 avgCT by 1 SD while proportionally shrinking the others to maintain the 100% total. From these 13 artificial samples, each with a mean age and DNAm-profile, we calculated the clocks’ DNAmAges. AgeAccel measures were obtained by subtracting the mean age of the artificial sample from its DNAmAge. (C) The same analysis as (A), now performed in the in silico-generated artificial cell mixtures

This analysis showed that increasing naïve CD4 or CD8 T cells by 1 SD led to negative AgeAccels for Hannum, PhenoAge, and DunedinPACE (effect sizes ranging from − 2.1 to -1.0 units). GrimAge showed this association for naïve CD4 T cells (effect size = -1.2) but not naïve CD8 T cells (effect size = -0.1), which may explain its weak associations with the naïve-to-memory T cells ratio (PC2, Fig. 4 and 55). Additionally, these four clocks demonstrated positive AgeAccel in response to an increase in neutrophils (effect sizes ranging from 1.1 to 1.8). Finally, we observed a negative AgeAccel for PhenoAge, GrimAge, and DunedinPACE in response to increased NK cells (effect sizes ranging from − 0.6 to -1.3). These patterns held when repeating the analysis with naïve CD8 T cells held constant (Fig. S9), indicating that correlation between cell fractions does not substantially affect the results.

To validate these findings in an independent dataset, we downloaded an external dataset of sorted blood cell types (GSE167998), and generated in silico artificial cell mixtures (Fig. 6B). We again started by simulating an average cell type composition, then generated additional artificial mixtures by raising one fraction by 1 SD while proportionally shrinking the others, and analyzed the effects of these changes on the clocks’ AgeAccels (Fig. 6C). This in silico experiment allowed us to test the direct effect of each cell fraction’s abundance on AgeAccels, suggesting that any found differences in AgeAccel are driven by cell composition. Despite the different datasets and methods, this analysis yielded similar effects as before, namely that naïve CD4 and CD8 T cells are negatively correlated with AgeAccel for Hannum, PhenoAge, GrimAge, and DunedinPACE (effect sizes ranging from − 0.4 to -1.7 years), while neutrophils are positively correlated (effect sizes ranging from 0.5 to 2.2). In addition, NK cells were negatively associated with GrimAge and DunedinPACE AgeAccels (effect size = -0.7, and − 0.6 years, respectively). Horvath and Zhang AgeAccels were once again much less affected by cell composition (effect sizes across all cell types ranging from − 0.5 to 0.3 years).

We conclude that the AgeAccels of Hannum and the three 2nd -generation clocks evaluated (PhenoAge, GrimAge, and DunedinPACE) are indeed driven by cell composition, especially by neutrophils and naïve T cells. The AgeAccels of the 1st -generation clocks Horvath and Zhang were much less influenced by shifts in cell composition, with only minor effects observed for naïve T cells.

Cell composition and the association of epigenetic clocks with incident health outcomes

Given the contribution of cell composition to AgeAccel, we evaluated the extent by which associations of AgeAccel with health outcomes are dependent on cell composition. Therefore, we tested the association of cell composition PCs and epigenetic clocks (both separately and in combination) with 176 incident health outcomes, including all-cause mortality, in 18,859 participants of the Generation Scotland cohort by adapting a previously published analysis [21].

We found 12 outcomes that associated with cell composition (p_Bonferroni < 0.05), including mortality, asthma, and non-Hodgkin lymphoma (Fig. 7A). Zooming in on the 11 individual cell composition PCs (Fig. 7B), mortality was associated with PC1 (neutrophils), PC3 (Treg/CD4Tmem), PC8 (Eos/Mono), and PC9/11 (Mixed). Notably, PC2 (T cell naïve/memory ratio) was not associated with mortality, despite its strong association with calendar age and DNAmAge. Another finding included that PC5 (Bnv/Eos), PC7 (CD4tnv/CD8Tnv), and PC8 (Eos/Mono) were associated with four outcomes, namely hyper-/hypothyroidism, non-Hodgkin lymphoma, asthma, and nasal polyps. For asthma, associations PCs involving neutrophils, eosinophils and basophils stood out, in line with the know pathophysiology of the disease [30, 31].

Fig. 7.

Effect of cell composition and epigenetic clocks on incident health outcomes. (A) Boxplots showing the p-values of the effect of cell composition (all PCs) on 176 incident health outcomes measured in Generation Scotland [21]. The 12 outcomes for which the p-value is lower than 0.05/176 (Bonferroni correction) are labeled. (B) Heatmap showing the Z-score of each individual cell composition PC on the 12 health outcomes identified in subfigure A. Blue, red, and white cells indicate positive, negative, and no association, respectively. Cells are labeled based on the nominal p-value of their association (*: p < 0.01, **: p < 0.001, ***: p < 0.0001). (C) Scatterplot of the hazard ratio (HR) of each epigenetic clock included in this study for each of the 176 health outcomes, with or without adjustment for cell composition PCs. The 12 health outcomes identified in subfigure A are labeled. The fitted regression line is shown for each clock, as well as the p-value testing whether the line differs from x = y (intercept 0, slope 1)

Finally, we evaluated the extent to which the association of AgeAccel with the 176 incident health outcomes was dependent on cell composition. For all 176 incident health outcomes, we fitted both a base model including only AgeAccel and demographic covariates, and a full model also including cell composition PCs, and tested whether the AgeAccel hazard ratios (HR) were different between the models. Across all health outcomes, adjustment for cell composition significantly attenuated the association of AgeAccels for 4 out of 6 clocks but at 4%, 6%, 7% and 2% for Hannum, Horvath, PhenoAge, and DunedinPACE, respectively, the effect was relatively minor (Fig. 7C). For mortality, the attenuation ranged from 1% for Horvath to 6% for PhenoAge (Table 2).

Table 2.

Association of AgeAccels with mortality, raw and adjusted for cell composition

Epigenetic clock	Raw		Adjusted
	HR (95% CI)	p-value	HR (95% CI)	p-value
Hannum	1.17 (1.11–1.23)	2.5*10− 10	1.12 (1.06–1.18)	2.6*10− 5
Horvath	1.13 (1.08–1.20)	3.5*10− 6	1.12 (1.07–1.19)	1.3*10− 5
Zhang	1.07 (1.01–1.13)	3.1*10− 2	1.05 (0.99–1.11)	1.0*10− 1
PhenoAge	1.32 (1.25–1.39)	9.1*10− 27	1.26 (1.19–1.33)	5.7*10− 17
GrimAge	1.57 (1.49–1.65)	5.0*10− 67	1.54 (1.46–1.63)	3.7*10− 50
DunedinPACE	1.46 (1.36–1.56)	1.6*10− 28	1.41 (1.32–1.52)	3.1*10− 22

Raw and cell composition-adjusted hazard ratios (HR) with 95% confidence interval (CI) and p-values are shown for each clock’s AgeAccel. HRs are calculated per 1 SD increase in Age Accel

Discussion

In this study, we show how collinearity complicates the analysis of cell fractions and present a PC-based approach to address these issues. Subsequently, we use our PC-approach to show that AgeAccel has markedly different associations with cell fractions as compared with calendar age and DNAmAge. We also show that the AgeAccels of different clocks diverge in their associations with cell fractions. We use artificial cell mixtures to show that changes in cell fractions likely drive the AgeAccel of epigenetic clocks rather than merely being correlated. Finally, we show that cell composition associates with health outcomes and that the association of AgeAccels with incident health outcomes are slightly attenuated after adjustment for cell composition.

Our finding that calendar age and DNAmAge (irrespective of clock) are associated with a shift from naïve to memory T cell phenotypes is consistent with literature [32] and our previous work [12]. In contrast, the AgeAccels of Hannum, PhenoAge, GrimAge, and DunedinPACE show a prominent association with neutrophils and greatly diminished association with T cells. For example, 9.5% of DunedinPACE’s AgeAccel can be explained by PC1 (representing neutrophils) and only 0.6% by PC2 (representing the naïve-to-memory T cell ratio). These findings are in line with a recent two-sample Mendelian randomization study of low-resolution cell counts implying a causal effect of neutrophil count on epigenetic clocks [33]. Recent causality-enriched clocks further suggest that clocks’ ability to capture health decline correlates with their association with cell composition [34]. Of these clocks, DamAge, capturing damaging age-associated methylation changes, showed strong associations with cell composition (variance explained = 19%, Additional File 1: Fig. S10), especially with PC1, which represents neutrophils. Conversely, AdaptAge (trained to mark protective changes) and CauseAge (with both damaging and protective changes) showed weak associations between AgeAccel and cell composition (variance explained = 8% and 4%, respectively, Additional File 1: Fig. S10).

The difference in associations of age/DNAmAge and AgeAccel indicate that a 1-year advancement in calendar age has a different biological interpretation than a 1-year AgeAccel as estimated by these clocks, at least in the context of cell composition. These results have implications for the interpretation of epigenetic clocks as descriptors of “biological age” [35]. For this interpretation to be fully valid, a 1-year advancement in calendar age would be expected to represent the same underlying biology as a 1-year AgeAccel (Fig. 8). In this study, we show that this is not the case for associations with cell composition. We speculate that age and AgeAccel could diverge in more ways, meaning that some of the observed associations of AgeAccel with mortality and morbidity [36] may stem from processes other than biological aging.

Fig. 8.

Common interpretation of AgeAccel as “biological age”. This figure shows a subset of 50 participants from our dataset, with 3 people highlighted. Person A is 50 years old, while Person B is 60 years old. For both of these people, the Horvath-clock accurately estimates age. However, for person C, the clock overestimates age by 10 years. A common interpretation is that Person C is has a ‘biological age’ of 60 years, implying that they resemble a 60-year-old, despite being only 50 years old. This interpretation requires DNAmAge and AgeAccel to have similar biological associations, which we do not observe for blood cell composition in this study

The AgeAccels of Horvath and Zhang show only very weak associations with cell composition. This can have multiple explanations. The one most commonly put forward is that these clocks’ AgeAccels capture cell-intrinsic aging information. This interpretation is supported by the fact that clocks can track calendar age even in isolated cell types [37, 38] and the fact that reprogramming into induced pluripotent stem cells (iPSCs) resets the clocks to zero [39]. However, since we only considered the 12 cell types included in the EpiDISH panel, the effects of other age-related immune cell types, such as effector, terminal effector, and exhausted T cell phenotypes, remains unknown [40]. For example, the most widely-described aspect of immunosenescence is remodeling of the CD8 T cell compartment, including outgrowth of terminal effector phenotypes, altered expression of cell surface markers such as CD28, CD57, and KLRG1, and increased cytotoxicity [41]. Since none of these changes can currently be captured by cell prediction algorithms, it remains unclear to what extent these changes are associated with DNAmAge and AgeAccel of the various clocks. Expansion of the cell type panel to include more age-related fractions will lead to more sophisticated insights, and may also expand the proportion of variance of clocks that is captured by cell composition. Additionally, our current work only applies to whole blood samples, since EpiDISH and most epigenetic clocks are trained on whole blood samples. The analysis of other tissues and sample types (including peripheral blood mononuclear cell samples which exclude granulocytes) is an important area of future study.

Since AgeAccel is in essence an error made by an age prediction model, part of its variation must be random noise, in particular when considering 1st generation clocks. The authors of the Zhang-clock noted that the more accurate a clock is at predicting calendar age, the less biologically relevant its AgeAccel becomes [4]. This poses the question if it is desirable for a biomarker of biological age to correlate strongly with calendar age. We observed that clocks whose AgeAccel is highly explainable by cell composition (Hannum, PhenoAge, GrimAge, DunedinPACE) have stronger associations with health outcomes such as mortality compared to Horvath and Zhang, which were less affected by cell composition, which matches the literature [35, 42]. This begs the question if a cell-intrinsic clock is desirable as an aging biomarker. Epigenetic association studies are often adjusted for cell composition effects, but changes in cell composition could plausibly be a process through which an aging signal manifests in DNAm, for instance by tracking chronic inflammation, which is a hallmark of aging [43].

We observed that AgeAccels of 2nd generation clocks were strongly associated with mortality and with neutrophils. Indeed, elevated neutrophil counts are associated with mortality, frailty, and multiple diseases [44–46]. At the same time, neutrophil counts do not change with chronological age. So, neutrophil count may be an example of a biomarker that is independent of the biological aging process, and yet is captured by epigenetic clocks to strengthen their association with health outcomes. Here, it should be noted that our analysis indicated that cell composition explains only a minor proportion (< 10%) of the associations of AgeAccels with incident health outcomes.

Methodologically, we analyze cell composition using a PC-based approach, since we show that univariable and multivariable models give ambiguous results due to collinearity. Our results show that methods that allow features to vary independently of one another can effectively address these challenges. We present two of such examples in this study: (1) a PC-based approach, or (2) increasing one fraction while proportionally reducing all others. A benefit of the PC-based method is that internal correlation structures between fractions are preserved, and we recommend this method when investigating the effects of cell fractions on a health outcome. While regressing on all available PCs is preferable in order to not miss any effects, testing single PCs only is also a valid option, since PCs are orthogonal and thus the estimated effect sizes are fully independent. For smaller data sets, newly calculated PCs may be unstable and difficult to interpret. Therefore we provided the option to project the PCs obtained in this study onto new datasets (Additional File 2), with the caveat that PC interpretation might be different in cohorts that differ greatly from ours, e.g. only including a narrow age range. A limitation of our method is that PCs must be manually labeled to assign biological meaning to results. Here, we label PCs based on their correlations with individual fractions. When adopting a PC-based method, it is critical to be transparent about the PC labeling process.

Conclusions

We conclude that calendar age primarily associates with changes in T cell fractions while the AgeAccels of epigenetic clocks are either majorly driven by neutrophils or mostly unaffected by cell composition, depending on the clock. Cell composition explains a substantial proportion of variation in AgeAccels, and has a modest effect on the association between AgeAccels and health outcomes. We expect that our results will aid in the interpretation of previous and future work in the field of epigenetic clocks.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

DNAmAge

DNA methylation age

AgeAccel

Age acceleration

PCA

Principal component analysis

PC

Principal component

Treg

Regulatory T cell

NK cell

Natural killer cell

Author contributions

Conceptualization: BTH. Methodology: THJ, EWZ, BTH. Formal analysis: THJ, AR, REM. Resources: BIOS Consortium. Writing: THJ, EWZ, BTH. Visualization: THJ. Supervision: EWZ, BTH. All authors have read and approved the final manuscript.

Funding

This research was supported by US NIH grant R01AG066887. The DNAm measurement was funded by the BBMRI-NL consortium (a research infrastructure financed by the Dutch government, NWO grants 184.021.007 and 184.033.111). Generation Scotland was funded by a grant from the Chief Scientist Office of the Scottish Government Health Directorates (CZD/16/6) and the Scottish Funding Council (HR03006). Genotyping and DNA methylation profiling of the GS samples was carried out by the Genetics Core Laboratory at the Edinburgh Clinical Research Facility, University of Edinburgh, Scotland, and was funded by the Medical Research Council UK and the Wellcome Trust (Wellcome Trust Strategic Award ‘STratifying Resilience and Depression Longitudinally’ Reference 104036/Z/14/Z). DNA methylation profiling was also funded in part by the Wellcome Trust (Reference 220857/Z/20/Z). REM is supported by UKRI1940.

Data availability

The main dataset of 4058 whole blood samples is available from the European Genome-Phenome Archive (EGAC00001000277) [47]. The external dataset of 56 sorted blood cell samples is available from Gene Expression Omnibus (GEO) under accession number GSE167998. According to the terms of consent for Generation Scotland participants, access to data must be reviewed by the Generation Scotland Access Committee. Applications should be made to genscot@ed.ac.uk and normally take up to six weeks for approval. Further details can be found at https://genscot.ed.ac.uk/for-researchers/access/. All analyses were performed using R version 4.3.1. For all packages installed from CRAN or Bioconductor, the standard version for this R-version was used. All scripts written for this study can be found on GitHub: https://github.com/ThomasJonkmanLUMC/Cellcounts_clocks [48]. Additionally, this repository contains integrated HTML-reports, generated by R Markdown, including both R-code and accompanying output and figures to ensure maximal transparency and reproducibility.

Declarations

Ethics approval and consent to participate

The study was approved by the institutional review boards of the participating centers (CODAM, Medical Ethical Committee of the Maastricht University; LL, Ethics Committee of the University Medical Centre Groningen; LLS, Ethical Committee of the Leiden University Medical Center; PAN, Institutional Review Board of the University Medical Centre Utrecht; NTR, Central Ethics Committee on Research Involving Human Subjects of the VU University Medical Centre; RS, Institutional Review Board (Medical Ethics Committee) of the Erasmus Medical Center). All participants have given written informed consent, and the experimental methods comply with the Helsinki Declaration. All components of Generation Scotland received ethical approval from the NHS Tayside Committee on Medical Research Ethics (REC Reference Number: 05/S1401/89). GS has also been granted Research Tissue Bank status by the East of Scotland Research Ethics Service (REC Reference Number: 25-ES-0013). All participants provided broad and enduring written informed consent for biomedical research. This study was performed in accordance with the Helsinki declaration.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.