DNA methylation patterns reflect individual's lifestyle independent of obesity

2.1 Study population

The present analyses included participants of the LIFE-Adult study, a population-based cohort of European ancestry, focusing on lifestyle diseases.^{8, 9} The cohort comprised ∼10 000 adult subjects aged from 18 to 80 years (mean ± standard deviation [SD]: age = 57.4 ± 12.5 years, body mass index [BMI] = 27.3 ± 4.9 kg/m²) from the region of Leipzig, Germany. All participants underwent extensive phenotyping, including anthropometric measurements, social and lifestyle-behaviour questionnaires and blood parameters. For most subjects, ethylenediaminetetraacetic acid (EDTA) blood samples are available.⁸ All participants gave written informed consent to participate in the study, and procedures were approved by the University of Leipzig's ethics committee (registration number: 263-2009-14122009) and conducted according to the Declaration of Helsinki. Study participation, assessments and interviews were supervised and carried out by trained staff and under supervised standard operation procedures.⁸

2.2 Lifestyle score

We created a lifestyle score (LS) as the sum of four different subscores: diet, physical activity (PA), alcohol consumption and smoking.¹⁰ To calculate the scores, we included data from four self-reported questionnaires: (1) a German version of the Food Frequency Questionnaire,¹¹ (2) the Short-Form International Physical Activity Questionnaire,¹² (3) a questionnaire about smoking status and quantity, and (4) a questionnaire about daily alcohol consumption and frequency. The final LS ranged from 3 to 66 (mean ± SD: 27.19 ± 11.2), with low and high LS values representing a healthy and an unhealthy lifestyle, respectively. All subscores showed an inter-item correlation, as demonstrated by Cronbach's alpha statistic (α = 0.64). A detailed description of the individual scoring and an explanation of each subscore can be found in Supporting Information and Table S1. Subjects with any missing questionnaire items were completely excluded from further analyses to avoid potential effects caused by general noncompliance of those subjects. Similarly, participants with pre-existing diabetic conditions (HbA1c ≥ 6.5%)¹³ or missing BMI measures were also excluded from subsequent analyses. A total of 4107 subjects passed all criteria (Table 1).

2.3 Subset for genome-wide methylation and validation measurements

Based on the LS calculation, we stratified the cohort into two groups reporting the most healthy and unhealthy lifestyles by selecting the lowest and highest 5% (5th percentile LS ≤ 11; 95th percentile LS ≥ 48). Within these groups (N = 234), we found 140 subjects with and 94 subjects without obesity according to BMI criteria.¹⁴ Based on this and an equal age range (Figure S1B), we further selected 25 subjects without (BMI < 25 kg/m²) and 25 subjects with obesity (BMI > 30 kg/m²) among each subgroup (5th vs. 95th percentile) (total N = 100) for the genome-wide methylation discovery cohort and included all (with and without obesity) subjects with sufficient available DNA (N = 213) for validation analysis. Thus, both groups overlap in N = 100 samples and are therefore not independent.

2.4 Sample preparation

All samples were isolated, stored and maintained at the Leipzig Medical Biobank¹⁵ according to standard protocols. Briefly, blood samples were taken after an overnight fast (mean fasting duration 12.7 ± 1.7 h) during the individuals’ study visit and stored at 4°C–8°C until DNA isolation (within 48 h after blood withdrawal) on the Autopure LS platform (Qiagen, Germany) using chemistry by Qiagen and Stratec Molecular (Stratec, Germany). Genomic DNA samples were stored at -80°C prior to integrity control using gel electrophoresis and concentration measurements of double-stranded DNA using Quant-iT PicoGreen dsDNA (Invitrogen, ThermoFisher Scientific, Germany) and Quantus (Promega, Germany) technologies.

2.5 Genome-wide DNA methylation analysis

Five hundred nanograms of genomic DNA was taken for bisulphite conversion using an EZ DNA Methylation Gold Kit (Zymo Research, Netherlands). After quality control (QC), amplification and hybridisation on Illumina HumanMethylation850 Bead Chips (Illumina, Inc., San Diego, CA, USA), the Illumina iScan array scanner was used to quantify genome-wide DNA methylation levels at 850K CpG sites per sample at single-nucleotide resolution.

Raw data were first quality controlled using the QC report of the minfi R package (version 1.38.0).^{16, 17} Two samples that did not pass the badSampleCutoff of 10.5 were excluded during normalisation steps. Beta densities and control probes were within predicted specifications. Probes that did not pass the detection p-value (p_detect = .01) in more than 1% of all 98 samples were excluded from the analysis (17 375 probes). Cross-reactive probes (38 924 probes)¹⁸ and probes containing known single-nucleotide polymorphisms (SNPs) (29 383 probes) at the CpG site (DNA region where a cytosine nucleotide is followed by a guanine nucleotide) were also filtered out by applying the maxprobes (version 0.0.2) and minfi R packages, respectively. In addition, probes on sex chromosomes were removed from the analysis subset (19 627 probes), as sex represents a larger source of variation in our methylation data. In total, 760 550 probes remained for the analysis. β-Value generation and quantile normalisation were computed using the minfi R package^{16, 17} and adjusted for sex-specific batch effects (see Figure S1A). Furthermore, we analysed the cell type composition using the Houseman approach^19-21 adapted to EPIC arrays by Salas et al.²¹ Possible differences in cell type composition were (see Figure S2) analysed using Wilcoxon tests in R. We corrected β-values for cell type composition in an attempt to reduce noise,²¹ although none of the cell type populations differed strongly between the subgroups (low and high LS) (Figure S2A) and the low and high LS subgroups in individuals without and with obesity (Figure S2B).

Differential methylation analyses were performed between subjects with extremely healthy and unhealthy lifestyles (low vs. high LS/5th vs. 95th percentile) as well as between participants without (BMI < 25 kg/m²) and with obesity (BMI > 30 kg/m²) within each lifestyle subgroup. The established R package limma (version 3.48.3) was used to identify differentially methylated CpG sites.²² Data assignment to technical and biological information by principal component analysis using the R package SWAMP (version 1.5.1)²³ showed that array slides were the primary batch for which we adjusted accordingly (Figure S3). Differentially methylated positions (DMPs) describe differences in methylation levels of single CpG positions with an adj. p-value <.05. Differentially methylated regions (DMRs) were extracted by applying DMRcate (version 2.6.0),²⁴ which uses Gaussian kernel smoothing to find patterns of differential methylation independent of genomic annotation. Only DMRs with more than two CpG sites were reported. DMRs with a minimum smoothed FDR <5% were defined as differentially methylated. DMRs with a mean methylation difference >±2% were further annotated to CpG islands (CpG shores, CpG shelves and inter-CpG islands (CGIs)) and gene context-related regions (promoters, 5ʹ untranslated regions (UTRs), exons, introns, 3ʹUTRs and intergenic regions). Genomic annotation was performed using the annotatr R package (version 1.18.1)²⁵ with respect to multiple annotations. To elucidate putative drivers of blood DNA methylation, separate analyses with individual covariates (smoking, diet, PA, alcohol, BMI and age) were performed. Intersection analysis for covariate-specific effects on lifestyle DMRs was performed using the UpsetR (version 1.4.0) package.²⁶ EPIC raw data are available at the Leipzig Health Atlas under https://www.health-atlas.de/studies/57.

2.6 Methylation age and telomere length clocks

DNA methylation age (DNAmAge), corresponding DNAmAge acceleration differences according to Horvath's clock (I and II), Levine's clock and the telomere length clock were estimated using the R package methylclock (version 0.7.5).²⁷

2.7 KEGG pathway overrepresentation

Candidate genes identified by significant DMRs (minimum smoothed FDR <5%) characterising lifestyle-specific methylation differences (healthy vs. unhealthy living subjects) and differences between subjects without and with obesity were taken forward for a KEGG pathway overrepresentation test using clusterProfiler::enrichKEGG (version 3.18.1).²⁸ Enrichment p-values were adjusted using Benjamini–Hochberg correction, and FDR <5% was considered statistically significant.

2.8 Validation of candidate CpGs

We selected two top candidate DMPs (Tables 2 and S8) from our discovery cohort (high LS vs. low LS) for additional validation using bisulphite sequencing. Briefly, 300 ng of genomic DNA was bisulphite converted using the EpiTect Fast DNA Bisulfite Kit (Qiagen). After a whole genomic amplification using the EpiTect Whole Bisulfitome Kit (Qiagen), candidate regions were amplified and sequenced using the PyroMark Q24 platform and self-designed assays for retinoic acid receptor alpha (RARA) and F2R like thrombin or trypsin receptor 3 (F2RL3) candidate DMPs (Qiagen). Primer sequences are shown in Table S3. All analyses were performed in duplicate, including two non-template controls per sequence run.

2.9 Transcriptome data

Transcriptome data were available from Illumina HT-12 v4 Expression Bead Chips (Illumina) using whole blood RNA samples from the LIFE-Adult cohort as described elsewhere.^{8, 29} Data processing was performed using R/Bioconductor after extraction of all 47 231 gene-expression probes using Illumina GenomeStudio without background correction. Furthermore, expression values were log2 transformed and quantile normalised,^{30, 31} and batch effect correction was performed using an empirical Bayes method.³² Probes were excluded when expressed in less than 5% of the (subgroup-specific) samples (detected by Illumina GenomeStudio), still being associated with batch effects after Bonferroni correction or not mapping to a gene accordingly³³ (accessed on 4 April 2019). Additionally, gene probes without available annotation and genes on the X and Y chromosomes were removed to determine the effects introduced by sex. In summary, 20 114 valid gene-expression probes were identified corresponding to 14 687 single genes in the human genome (hg19). A three-step procedure was used to remove poor quality samples: (1) first, the number of detected gene-expression probes of a sample was required to be within ±3 interquartile ranges (IQR) from the median, (2) the Mahalanobis distance of several quality characteristics of each sample (signal of AmbionTM ERCC spike-in control probes, signal of biotin-control probes, signal of low-concentration control probes, signal of medium-concentration control probes, signal of mismatch control probes, signal of negative control probes and signal of perfect-match control probes)³⁴ had to be within median +3 × IQR, and (3) Euclidean distances of expression values as described³¹ had to be within 4 × IQR from the median. Overall, of the 3526 assayed samples, 107 samples were excluded for quality reasons. Raw transcriptome data are available at the Leipzig Health Atlas under https://www.health-atlas.de/studies/57.

2.10 Genotype data

For genotypes, 7838 LIFE-Adult participants were genotyped using the genome-wide SNP array Affymetrix Axiom CEU1 and the software Affymetrix Power Tools (version 1.20.6). QC of the genotyped data was performed following Affymetrix's data analysis guide³⁵ as previously described.³⁶ QC according to Affymetrix's data analysis guide included dish-QC (<0.82), sample call rate (<97%), sex mismatch, ambiguous relatedness (e.g., sample mix-up) and abnormalities of XY intensity plots (e.g., XXY samples filtered for gonosomal analyses). Genetic heterogeneity was evaluated with principal component analyses, and outliers (>6 SD in any of the first 10 principal components) were removed. The criteria call rate, parameters of cluster plot irregularities according to Affymetrix's recommendations, violation of Hardy–Weinberg equilibrium (p-value < 10^-6) in an exact test for autosomes (p-value < 10^-4), for chromosome X with females only³⁷ and batch association (p-value < 10^-7) were considered during SNP QC. Subsequently, 7669 samples and 532 676 SNPs were imputed on the reference 1000 Genomes Phase 3,³⁸ applying SHAPEIT³⁹ v2r900 (prephasing) and IMPUTE2 (version 2.3.2)^{39, 40} for genotype estimation. A specific genotype was assigned to a SNP if its corresponding genotype estimation featured a probability of ≥.8.⁴¹ In 2.5% of the cases, none of the genotypes exceeded that threshold, and the respective SNP was labelled ‘missing’ for that sample. SNPs whose ‘missing’ count over all samples exceeded the upper quartile +1.5 × IQR were removed, resulting in a total of 2830 SNPs.

2.11 matrixEQTL analysis

Among samples with significant DMP data (from DMRs healthy vs. unhealthy), additional gene-expression and SNP data were available for 48 samples. The R package matrixEQTL (version 2.3)⁴² was employed on all three pairs of datasets to identify cis effects (within a range of ±1 kb) between methylation and expression (eQTMs), methylation and SNPs (meQTLs) and expression and SNPs (eQTLs). All three comparisons were performed on the complete data (N = 48) and both subgroups with high LS (N = 23) and low LS (N = 25) separately. Since sex batch effects have been adjusted for in both expression and methylation data, small batch effects remained only for age and BMI. However, including both age and BMI as covariates into the matrixEQTL analysis did not change the overall result, which is why the final matrixEQTL analysis was run without considering any covariances.

2.12 Statistics

All statistical analyses were performed using R software (version 4.0.4).⁴³ After checking for normal distribution, Mann–Whitney U-test or Welch's t-test was applied to test for differences between the 5th and 95th percentiles as well as between lean and obese subgroups for the following phenotypes: BMI, age, HbA1c, waist-to-hip ratio (WHR), fasting plasma glucose and insulin, low-density lipoprotein (LDL), high-density lipoprotein (HDL), apolipoprotein A1 (Apo A1) and triglyceride serum levels. Welch's t-test was used to compare methylation differences measured as normalised β-values between low LS versus high LS for each top DMP, respectively. Using the Shapiro–Wilk test to prove the normal distribution of the bisulphite sequencing data, an independent Mann–Whitney U-test was applied to compare methylation differences within the validation cohort. Methylation levels between BMI categories were compared by applying two-way analysis of variance (ANOVA). Correlation analysis was performed using Spearman's correlation. All respective analyses were adequately corrected for multiple testing.

3.1 Self-reported lifestyle reflects obesity-specific phenotypes

We correlated the LS with BMI and WHR in 4107 LIFE-Adult participants (mean ± SD: age = 56 ± 13 years, BMI = 27.0 ± 4.7 kg/m², LS = 27.19 ± 11.02) (Table 1). LS was related to the obesogenic environment (Figure S4A,B) (all p-value < 1 × 10^-3) with significantly higher values in subjects with obesity (Figure S4C). We further demonstrated that all individual scores (diet, PA, smoking, alcohol and total LS) were mutually dependent (Figure 1A), which was particularly marked in the extreme subgroups (5th and 95th percentile, Figure 1B). Finally, our score showed simultaneous negative correlations (all p-value < 1 × 10^-15) to the protective lipid parameters HDL cholesterol and Apo A1, which were higher in healthy living subjects (Figure S5A,B).

The discovery cohort comprised 44 males and 56 females with a mean BMI of 28.6 ± 6.5 kg/m². The validation cohort (N = 213; 5th/95th percentile = 100/113; lean/obese = 131/82) included 84 males and 129 females with a mean BMI of 27.1 ± 6.2 kg/m² (Table S2). When comparing healthy versus unhealthy subgroups (5th vs. 95th percentile LS, healthy LS: N = 216, unhealthy LS: N = 207), nominal differences were observed regarding BMI distribution (Figure S5, Table 1) (N lean/overweight/obese: healthy LS—73/95/48, unhealthy LS—67/93/46; p = .051). Although HbA1c, fasting plasma glucose, fasting plasma insulin and LDL serum levels did not differ between the two groups (all p-value > .05, Table 1), strong differences were found for waist circumference (p-value < 3 × 10^-6), age, Apo A1 and triglyceride serum levels (all p-value < 1 × 10^-9) and especially HDL levels (p-value < 1 × 10^-15) and WHR (p-value < 2.2 × 10^-16). Subjects living an extremely healthy lifestyle were older (mean age difference = 6 years) with lower WHR (mean WHR difference = 0.08) and prominently lower lipid serum levels. However, as expected, all phenotypes differed significantly between subjects with and without obesity within healthy and unhealthy subgroups (all p-value < 2 × 10^-2), respectively, except for fasting plasma LDL. Regarding sex distribution, it is noticeable that females are overrepresented in the healthy subgroup, whereas males (Table 1) dominate the unhealthy subgroup.

3.2 DNA methylation signatures are related to individual's lifestyle

By comparing genome-wide blood DNA methylation patterns in subjects with healthy versus unhealthy lifestyles, we identified 4682 significant DMRs annotated to 4426 genes with a minimum smoothed FDR <5%, which included 220 DMRs with FDR <1 × 10^-4 (Table S4, Figure 2A). Among the significant DMRs, the mean methylation level differences ranged from -6.9% to 5.5%.

Given the rather subtle methylation changes for the majority of the DMRs, we introduced a mean methylation threshold of >2% to further narrow down the potential causal candidate DMRs. Among the 340 DMRs reaching this cut-off, 164 DMRs were hypermethylated (mean methylation difference ± SD: 2.6 ± 0.6%), whereas 176 DMRs were hypomethylated (mean methylation difference ± SD: -2.8 ± 0.8%) in healthy compared to unhealthy living individuals (Figure 2A, Table S4). Taking into account that a DMR can have more than one annotation, most DMRs (46%; counts relative to the number of DMRs) were located in CpG islands, followed by 45% located in CpG shores. In relation to gene regions, most DMRs are located in introns (59%), followed by exons (39%) (Figure 2B).

The top 15 hypomethylated and hypermethylated significant DMRs according to their mean methylation difference are presented in Table 3 with a DMR annotated to the glutamine-fructose-6-phosphate transaminase 2 (GFPT2) gene locus showing the strongest hypomethylation (mean methylation difference DMR = -6.9%). A DMR annotated to the glutamate rich 1 (ERICH1) gene showed the strongest hypermethylation (mean methylation difference DMR = 5.4%). Finally, using KEGG pathway analyses, we identified glutamatergic synapses as the most enriched pathway (adj. p-value < .01) followed by axon guidance, another brain-related pathway (adj. p-value < .05). Most of the nine enriched pathways (Table S5, Figure 2C) are related to various cancer types.

As demonstrated by the intersection plot (Figure 2D, Table S6), the majority of the DMRs (N = 1952) were driven by all four lifestyle subscores together (diet, PA, smoking and alcohol), followed by a combination of them together with BMI and age (N = 743). Obviously, BMI and age alone do not explain any of the identified DMRs. Although this did not indicate a prominent role of smoking (Figure 2D), given the nature of the LS, a comparison between participants with very healthy and very unhealthy lifestyles mirrors differences between nonsmokers and smokers. We therefore further adjusted the complete analyses for smoking as a covariate, which resulted in 629 identified DMRs with a minimum smoothed FDR <5%. Among them, the most significant DMR is located within a CpG island of the ring finger protein 39 (RNF39) locus (Table S7).

3.3 Lifestyle-derived DMPs correlate with metabolic traits related to obesity

DMP-specific analysis comparing subjects with healthy and unhealthy lifestyles identified 145 significant DMPs (adj. p-value < 5%) (Figure 3A, Table S8). A total of 26 DMPs passed a mean methylation change |(log fold change FC])| ≥5%, 14 of which were hypermethylated log FC ≥5% (0.07 ± 0.04), while 12 showed hypomethylation log FC ≤-5% (-0.06 ± 0.01). Of these, 19 DMPs were clearly assigned to a specific gene. However, when considering significance levels (adj. p-value < .05) as well as a mean methylation change ≥5%, the strongest effects were observed for aryl hydrocarbon receptor repressor (AHRR), F2RL3, RARA and serine protease 23 (PRSS23) (Figure 3B, all p-value < 1 × 10^-11, Table S8), all of which were hypomethylated under unhealthy conditions. Correlation analysis between the identified DMPs and obesity-related phenotypes (HbA1c, LDL, HDL, blood glucose, insulin, Apo A1, triglycerides, BMI, waist circumference, WHR and age) as well as the LS and its subscores can be found in Table S9. Two of the 20 included DMPs showed a significant correlation with WHR even after Bonferroni correction, although no SNPs within the genomic region have been shown to be associated with BMI-adjusted WHR in a previous GWAS⁴⁴ (GWAS catalogue, accessed June 2021). Among the identified DMPs, the strongest correlation with WHR was found for vasohibin 1 (VASH1) (p-value = 4.2 × 10^-5, r = -0.4), which is in line with an association with HbA1c (%) (p-value = .03, r = -0.21), triglycerides (mmol/L) (p-value = .046, r = -0.2) and waist circumference (cm) (p-value = .04, r = -0.2), although only the association with WHR was sustained after correction for multiple testing. Finally, two DMPs of F2RL3, one of our selected top hits, showed marginal correlations to LDL serum levels (all p-value < .03, r = -0.22) and triglycerides (mmol/L) (all p-value < .07, r = -0.27) and one DMP in addition to HbA1c (%) (p-value < .01, r = -0.3). The selected top DMPs for F2RL3 showed additional correlations to HDL (mmol/L) (p-value = .03, r = 0.23) and Apo A1 (g/L) (p-value = .03, r = 0.23), albeit not withstanding corrections for multiple testing.

In summary, 13 of the 20 selected DMPs showed marginal correlations with triglycerides (all p-value < .046), eight with HDL (all p-value < .047) and seven with LDL (all p-value < .033) serum levels (Table S9).

3.4 Bisulphite sequencing within RARA and F2RL3 loci supports findings from the discovery stage

Based on these findings and further supported by previously reported data,^45-47 we selected two DMPs (F2RL3:cg03636183 and RARA:cg17739917) for validation (N = 213) using bisulphite sequencing and demonstrated directionally consistent differences between subjects with healthy versus unhealthy lifestyles (Figure 3C,D; both p-value < .001, mean methylation difference: RARA CpG = 8.37%, F2RL3 CpG2 = 13.87%). Furthermore, the surrounding CpG position confirmed this effect, with all p-value <.01 and a mean methylation difference of >5.59% (mean methylation difference: RARA CpG2 = 11.99%, RARA CpG3 = 12.58%, F2RL3 CpG1 = 5.59%, F2RL2 CpG3 = 13.2%) (Figure 3C,D). The validation cohort used here did not differ significantly from the discovery cohort regarding age, BMI and sex distribution (Table 1). Spearman's correlation analysis showed a significant correlation between the methylation levels observed in the genome-wide methylation analysis and bisulphite sequencing (RARA: Spearman's r = 0.45, F2RL3: Spearman's r = 0.53, both p < 7.9 × 10^-5) (Figure S5C,D).

3.5 Obesity-specific methylation marks

Driven by the comparable distribution of subjects with and without obesity between the very healthy and unhealthy lifestyle groups, we aimed to identify lifestyle-independent obesity-related methylation marks. Therefore, the blood methylation patterns of subjects with obesity (N = 25) were compared with those of subjects without obesity (N = 25) within each lifestyle group separately. Interestingly, whereas approximately 1572 DMRs annotated to 1599 different genes were identified in healthy subjects, only 85 DMRs annotated to 101 genes were detected in subjects living an unhealthy lifestyle (Tables S10 and S11) with a minimum smoothed FDR <5%. This further included 10 identical annotations among the PAX6 and HOXA9-10 gene clusters, already known candidates regarding obesity and related comorbidities. However, at CpG levels, no DMPs were sustained after correction for multiple testing (data not shown). Nevertheless, KEGG pathway analysis for the healthy subgroup indicated eight enriched pathways, among them GABAergic synapse, dilated cardiomyopathy and calcium signalling (Figure 4, Table S12), whereas for the unhealthy subgroup, only antigen processing and presentation were enriched (not shown).

3.6 Methylation age

We observed the strongest association (p-value < 1 × 10^-10, R² = 0.37, Figure 3E) between mAge and subjects’ chronological age within the discovery cohort for Horvath's clock II, which was compared to Horvath's clock I, which was additionally trained on 850K EPIC arrays (Figure S6A–C). Only marginal (p-value = .01) differences in DNAmAge acceleration were observed when comparing individuals with healthy versus unhealthy lifestyles, which was similar to comparing never smokers with previous or current smokers (Figure S6D,F). No difference was observed between subjects with and without obesity (Figure S6E). We further observed a strong negative association between the telomere length clock and chronological age (p-value < 1 × 10^-8, R²  =  -0.32, Figure 3E). Interestingly, both clocks showed an additional linear association with WHR within our discovery cohort (all p-value < 1 × 10^-4, Figure 3E).

3.7 Underlying genetic predispositions and effects on mRNA levels in blood

Driven by the small overlapping sample size (N = 48) and only marginal genetic variation in close proximity (±1 kb) to the identified target DMRs (healthy vs. unhealthy lifestyle), we could not identify any meQTLs or eQTLs. However, we found associations between the methylation levels of eight DMPs with target mRNA expression levels (Table S13) in the combined discovery group (all individuals with healthy and unhealthy lifestyles). Among them, only two eQTMs annotated to the alanyl aminopeptidase (ANPEP) locus were sustained after correction for multiple testing (matrix FDR = 0.03; Table S13, Figure S7). Four eQTMs were detected in subjects with healthy lifestyle and eight with unhealthy lifestyle; among them, one of our candidate DMP of F2RL3 was also detected in healthy subjects; however, none maintained after correction for multiple testing.