2.1. Ethics statement

All components of Generation Scotland received ethical approval from the NHS Tayside Committee on Medical Research Ethics (REC Reference Number: 05/S1401/89). Generation Scotland has also been granted Research Tissue Bank status by the East of Scotland Research Ethics Service (REC Reference Number: 20-ES-0021), providing generic ethical approval for a wide range of uses within medical research. Written informed consent was obtained from all participants. This study was performed in accordance with the Helsinki declaration.

2.2. Generation Scotland cohort

Generation Scotland, or GS, is a large family-structured cohort study that consists of 24,000 individuals from across Scotland. Participants were identified via Community Health Index numbers, with the support of Scottish Practices and Professionals Involved in Research. The initial phase of recruitment (2006 to 2010) focussed on the Glasgow and Tayside regions of Scotland and was later extended to Ayrshire, Arran, and the Northeast of Scotland. Individuals must have been aged between 35 and 65 years, had ≥1 first-degree relative and ≥1 full sibling. The age range was later broadened to 18 to 65 years. Family members of probands were also invited to partake in the study. In total, 23,960 individuals were recruited, which encompassed 6,665 probands, 16,007 family members, and 1,288 individuals who self-volunteered without invitation. There were 5,573 families with a mean size of 4 members and 1,400 participants without relatives. The median age at baseline was 47 years and the sample was 59% female [12,13]. Detailed health and lifestyle information were collected via questionnaires at the study baseline alongside venepuncture to obtain whole blood samples from which DNAm was assayed. This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline (S1 STROBE Checklist).

The present study does not have a registered prospective protocol. An unpublished, informal analysis plan was made and discussed among study authors prior to the implementation of statistical analyses (August 2022). There were no significant changes to the analysis plan following informal review among the study authors, with the exception of pathway enrichment and outlier sensitivity analyses following peer review.

2.3. Preparation of DNA methylation data

Whole-blood DNAm was measured using the Illumina Infinium MethylationEPIC array. DNAm profiling of the GS samples was carried out by the Genetics Core Laboratory at the Edinburgh Clinical Research Facility, Edinburgh, Scotland. Methylation typing was performed in 3 distinct sets. Quality control steps are detailed in full in S1 Text. Following quality control, there were 5,087, 4,450, and 8,876 individuals within Sets 1, 2, and 3, respectively. Set 1 contained related individuals. Set 2 consisted of individuals who were unrelated to each other and those in Set 1. Set 3 consisted of related individuals, and individuals related to those in Sets 1 and 2. The sets were combined and dasen normalisation was performed across all individuals [14]. Linear regression models were used to adjust methylation M-values for chronological age, sex, and experimental batch (factor with 121 levels, i.e., individuals were assayed across 121 unique batches). Residualised M-values were taken forward for analyses. In total, 752,722 probes and 18,413 individuals passed quality control criteria and were considered as a single analytical sample in our analyses.

2.4. Preparation of disease phenotypes

Nineteen common disease states were considered across prevalence and incidence analyses: (i) Alzheimer’s dementia (AD); (ii) breast cancer; (iii) chronic kidney disease (CKD); (iv) chronic neck and/or back pain; (v) chronic obstructive pulmonary disease (COPD); (vi) colorectal cancer; (vii) Coronavirus Disease 2019 (COVID-19) severity (requiring hospitalisation); (viii) inflammatory bowel disease (IBD); (ix) ischemic heart disease; (x) liver cirrhosis; (xi) long COVID; (xii) lung cancer; (xiii) osteoarthritis; (xiv) ovarian cancer; (xv) Parkinson’s disease; (xvi) prostate cancer; (xvii) rheumatoid arthritis; (xviii) stroke; and (xix) type 2 diabetes. Outcomes were selected if they were present among the 10 leading causes of death in high-income countries, the 10 leading causes of disease burden (disease-adjusted life years (DALYs)) in high-income countries or self-reported conditions at the baseline [15–17]. Depression was not considered as it is included in an ongoing meta-analysis EWAS. Although asthma can occur at any age, it has a higher prevalence among children aged 0 to 17 years than in adults. It was therefore excluded from the present analyses that used an adult sample with a broad age profile [18].

Self-report data were used for 12 disease states in cross-sectional analyses of disease prevalence. Self-reported parental history of AD was used a proxy variable for AD. Analyses on self-reported parental history of AD were restricted to participants who were >45 years at baseline. This ensured that only participants whose parents were likely old enough at baseline to be at risk of AD were considered (i.e., >65 years). The CKD Epidemiology Collaboration, or CKD-EPI, equation was implemented to estimate glomerular filtration rate (eGFR) at baseline. Individuals with an eGFR <60 ml/min/1.73 m² were deemed to have CKD [19]. Therefore, 14 disease phenotypes were considered in prevalent analyses.

All 19 phenotypes were included in longitudinal analyses via linkage to electronic health records (with the exception of self-reported long COVID). The primary and secondary care codes used to define incident phenotypes are available in S1 Appendix. Prevalent cases from the study baseline were excluded for these analyses as were those where record linkage provided evidence of a diagnosis prior to baseline. Therefore, incident cases included those diagnosed after baseline who had died and those who received a diagnosis and remained alive. Controls were censored if they were free of a diagnosis at the time of death or at the end of the follow-up period. Further information on the preprocessing of incident phenotypes, including COVID phenotypes, is available in S2 Text.

2.5. Epigenome-wide association studies on prevalent disease

First, logistic regression models were used to adjust prevalent phenotypes for chronological age and sex, with the exception of breast cancer and prostate cancer, which were adjusted for age after restricting the cohort to females and males, respectively. Second, linear regression models were used for EWAS via the OSCA (OmicS-data-based Complex trait Analysis) software [20]. Residuals from logistic regression models were entered as the dependent variable and age-, sex-, and batch-adjusted CpG M-values represented the independent variable. This strategy was employed to reduce computational burden. A Bonferroni significance threshold was set at p < 2.6 × 10⁻⁹ (= 3.6 × 10⁻⁸/14 phenotypes) [21]. Two models with different covariate strategies were employed, as described below:

1. 1. Basic model: Phenotype and CpG M-values, processed as described above, and 5 Houseman-estimated white blood cell (WBC) proportions as fixed effect covariates [22]. Six cell types are estimated from the Houseman method. However, their proportions sum to 100%. Therefore, the percentage of granulocytes was not included in this analysis given that it is collinear with the other 5 cell types. The basic model was as follows:

Phenotype (residuals) ~ CpG M-values (residuals) + 5 methylation-predicted WBC proportions.

1. 2. Fully adjusted model: additional adjustments for 5 common lifestyle factors, which were alcohol consumption, body mass index, deprivation index (Scottish Index of Multiple Deprivation), methylation-based smoking score (EpiSmokEr) [23], and years of education. Body mass index was log transformed prior to analysis. Furthermore, multidimensional scaling was applied to GS genotype data to obtain an estimate of population structure. The first 20 genetic principal components were extracted and included in our analytical models. The fully adjusted model was as follows:

Phenotype (residuals) ~ CpG M-values (residuals) + 5 methylation-predicted WBC proportions + alcohol consumption (units/week) + log(body mass index (kg/m²)) + deprivation index (Scottish Index of Multiple Deprivation) + education (an 11-category ordinal variable) + methylation-based smoking score (EpiSmokEr) + 20 genetic PCs (population structure).

Results from basic and fully adjusted models are presented within the main text. Both models are included to assess the effects of lifestyle factors on associations between methylation sites and common disease states. Some covariates may be more appropriate for one disease state over another (e.g., body mass index for type 2 diabetes versus cigarette smoking for COPD). However, all 5 risk factors are included in an effort to capture the most common environmental and lifestyle risk factors across a broad range of disparate conditions. We do not further present unadjusted analyses (i.e., using DNAm data that are not adjusted for age, sex, and batch effects) given the strong, possible confounding effects of age and technical variation on associations between CpG methylation and age-related disease states. We also did not initially adjust for family structure in our models. However, we later ran a series of sensitivity analyses (outlined in Section 3.5), including adjustment for relatedness between participants.

2.6. Epigenome-wide association studies on incident disease

First, Cox proportional hazards models were used to adjust incident phenotypes for age at baseline and sex (17/19 phenotypes). Only age was included for breast, ovarian, and prostate cancer. Time-to-onset for the disease, or censoring, was the survival outcome in Cox proportional hazards models. Only individuals with an age at event or censoring ≥65 years were considered for AD. As outlined above, controls were censored at the time of death or at the end of the follow-up period. Logistic regression models were used to adjust 2 remaining COVID phenotypes prior to EWAS analyses. Cox models were not employed for COVID phenotypes owing to the limited differences in time-to-event data between individuals with positive COVID diagnoses. Whereas DNAm was corrected for age at baseline (as well as sex and batch), COVID phenotypes were adjusted for sex and age at COVID testing or diagnosis. Here, age at COVID testing or diagnosis was considered given the variation in time elapsed between baseline visits (between 2006 and 2011) and the onset of the COVID pandemic. Second, martingale residuals or logistic regression residuals were extracted and included as dependent variables in OSCA. A Bonferroni-corrected significance threshold was set at p < 1.9 × 10⁻⁹ (= 3.6 × 10⁻⁸/19 phenotypes). Basic and fully adjusted models were employed, as described in the previous section. Methods for sensitivity EWAS analyses are detailed under S3 Text.

2.7. Pathway enrichment analyses

Enrichment was assessed among Kyoto Encyclopaedia of Genes and Genomes (KEGG) pathways and Gene Ontology (GO) terms using the gometh() function in the R package missMethyl [24]. This function models the relationship between the number of probes per gene and the probability of being selected, accounting for the selection bias associated with probe-dense genes. The top 100 CpGs (i.e., smallest EWAS p-values) from each fully adjusted model were included as input features. There were 33 such models for consideration (14 prevalent and 19 incident models). Pathways with an FDR-adjusted p-value < 0.05 were deemed significant.

2.8. Structured literature review on blood-based EWAS of common disease

MEDLINE, Embase (Ovid interface, 1980 onwards), Web of Science (core collection, Thomson Reuters), and preprint servers were searched to identify relevant articles indexed as of March 27, 2023. The initial search dates were between August 1, 2022 and August 31, 2022, and later updated and performed again on March 27, 2023. We used the following search terms or their synonyms appropriate to each database: (“blood”.mp OR “whole blood”.mp OR “peripheral blood.mp”) AND (“EWAS” OR exp “epigenome-wide*” / OR exp “epigenome-wide association” /) AND (the disease of interest, e.g., “COPD” OR “chronic obstructive pulmonary disease”). The search strategy returned approximately unique 2,000 articles, of which 54 passed inclusion criteria. Inclusion criteria were as follows: (i) original research article; (ii) EWAS performed with blood DNAm; (iii) there were at least 20 individuals in each comparison group (i.e., cases and controls); and (iv) the study examined at least one of the 19 common disease states outlined in our study.

Here, we make an important distinction between systematic reviews and our structured literature review. The structured search of the literature was intended to identify appropriate studies for look-up analyses using a predefined and agreed list of search terms. This is similar to systematic reviews in that search terms are used to systematically screen literature databases. However, our approach differed from a systematic review in that no original or meta-analyses were performed using data from the literature beyond a look-up analysis of CpGs identified in these studies. Unlike a systematic review, the approach also does not provide an estimate for a clinical question and rather summarises the current EWAS literature.

First, we wished to examine whether the CpG associations identified in our study had been previously described. A CpG site was declared as novel in our study if it was not previously reported at experiment-wise significance thresholds deemed by each of the 54 studies. Of note, these studies used different significance thresholds. Several studies did not make their full summary statistics available, which prohibited the use of a common significance threshold for look-up analyses. However, the studies also differed from one another with respect to methylation arrays, phenotype definitions, and covariate strategies. We focussed on unique CpGs rather than unique genomic locations. Look-up analyses were performed separately for each condition following our structured literature review. Second, we aimed to determine the level of agreement among studies that examined the same condition with respect to locus discovery. Here, our study was ignored as we were only interested in the previous literature for this analysis. A CpG site or its gene (if available) was considered to be replicated if it was reported as significant (at thresholds set by each study) in at least 2 studies that examined the same condition. While focusing on genes alone may neglect intergenic CpGs, the CpG-level and gene-level look-up analyses are included together in an effort to capture as much information as possible from disparate studies in the literature.

2.9. Colocalisation analyses

Colocalisation analyses required GWAS summary statistics for CpG sites (i.e., methylation Quantitative Trait Loci–mQTLs, trait 1) and for respective disease states (trait 2; [25–30]). The GoDMC mQTL resource represents the largest mQTL study to date in terms of sample size but only focused on 450k array sites [31]. Therefore, the GoDMC resource was utilised for sites that are common to the EPIC and 450k arrays. However, mQTL analyses were also conducted in GS due to the need to generate mQTL summary statistics for sites present on the EPIC array only (S4 Text). In instances where CpGs had associations in both GS and GoDMC, we used the following criteria to determine which dataset to retain: (i) the dataset must have >10 genetic variants available and (ii) if both datasets satisfy (i), then retain the dataset with the larger sample size. Of note, GS served as the replication cohort within the original GoDMC analyses. Effect sizes in GS and GoDMC showed correlation coefficients of 0.97 and 0.96 for cis and trans variants, respectively, in the original GoDMC publication [31]. We observed a similar coefficient of 0.97 between effect sizes for the subset of CpGs used in our colocalisation analyses. Therefore, there was likely little heterogeneity between the data sources used in our workflow.

The coloc.abf() function in the R package coloc was used to test for colocalisation and default parameters were applied (version 5.1.0) [32]. SNPs ±1 Mb surrounding each CpG site were extracted from mQTL datasets (i.e., GS or GoDMC, trait 1) and disease GWAS summary statistics (trait 2). The method tests for 5 mutually exclusive hypotheses, H0: there are no causal variants for either trait in the tested region; H1 and H2: causal variant for trait 1 and trait 2 only, respectively; H3: distinct causal variants for both traits; and H4: the traits share a causal variant. Posterior probabilities ≥95% for H4 provided strong evidence in favour of colocalisation.

3.1. Demographics and disease counts in Generation Scotland

The mean age of the sample was 47.5 years (n = 18,413, standard deviation (SD) = 14.9) and the sample was 58.8% female. Summary data for demographic variables are presented in Table 1. Additional data on covariates and disease counts are displayed in S1–S3 Tables. The number of self-reported cases for prevalent disease at baseline ranged from 34 participants with Parkinson’s disease to 5,296 with chronic neck and/or back pain, respectively (basic model). Further, the number of cases with incident disease since baseline (derived from health record linkage) ranged from 31 for severe COVID (hospitalisation from COVID-19 infection) to 1,886 for chronic neck and/or back pain. Associations between covariates and disease states are displayed in S4 and S5 Tables for prevalent and incident disease states, respectively (also available in S1 and S2 Figs).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Summary of demographic variables in the Generation Scotland cohort.

https://doi.org/10.1371/journal.pmed.1004247.t001

3.2. Epigenome-wide analyses of prevalent disease

We first tested for cross-sectional associations between blood CpG methylation and 14 disease states at the study baseline. There were 1,340 significant associations across 10 diseases in a basic model that adjusted for age, sex, and estimated blood cell proportions (p < 2.6 × 10⁻⁹; Fig 2A, S6 Table). Over 90% of these associations (n = 1,246) were attributed to type 2 diabetes (n = 703 associations, 52.5%), COPD (n = 301, 22.5%), and chronic pain (n = 242, 18.1%). Genomic inflation factors ranged from 0.8 to 1.6 across all basic models (S7 Table). Look-up analyses in the EWAS Catalog showed that 617/1,340 associations involve CpGs that were previously associated with common disease risk factors including body mass index, smoking, and alcohol consumption [33]. For clarity, we do not present summary statistics (i.e., 95% CIs and p-values) for all individual CpG associations in the main text given the large number of associations present in basic and fully adjusted models. However, these are made available in S6 and S8 Tables, respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Epigenome-wide association studies on 14 prevalent disease states in Generation Scotland.

(A) Diseases that had CpG associations in only the basic model or the fully adjusted model are shown in bold. Colorectal cancer was present in both the basic and fully adjusted model, but no CpGs were common to both models for this condition. (B). Ideogram showing 69 sites that were common to both the basic and fully adjusted models. These loci associated with 4 unique disease states. Full information is available in S8 Table. Image created using Biorender.com. CKD, chronic kidney disease; COPD, chronic obstructive pulmonary disease; CpG, cytosine-phosphate-guanine dinucleotide; WBC, white blood cells.

https://doi.org/10.1371/journal.pmed.1004247.g002

Next, we conducted a fully adjusted model that further accounted for 5 common lifestyle risk factors and population structure. The 5 risk factors were alcohol consumption, body mass index, deprivation (Scottish Index of Multiple Deprivation), a methylation-based proxy for tobacco smoking [23], and years of education. The fully adjusted model returned 78 associations across 8 disease states (p < 2.6 × 10⁻⁹; Fig 2B, S8 Table). Sixty-nine associations from the basic model were also present in the fully adjusted analysis. The 69 associations were spread across 4 disease states: CKD (n = 1); ischemic heart disease (n = 6); breast cancer (n = 10); and type 2 diabetes (n = 52). Genomic inflation factors ranged from 0.8 to 1.8 across all fully adjusted models and were 1.1, 1.8, 1.0, and 1.1 for CKD, ischemic heart disease, breast cancer, and type 2 diabetes, respectively (S7 Table). The significant findings included associations between self-reported history of breast cancer and hypomethylation within cg06072257 and cg06123699, which are located near UBIAD1 and TPRG1 on chromosomes 1 and 3, respectively (p = 6.5 × 10⁻¹⁰³ and p = 2.4 × 10⁻¹⁰¹, respectively). The site cg17944885 located between ZNF788 and ZNF20 on chromosome 19 associated with prevalent CKD (p = 1.7 × 10⁻¹²). Furthermore, CpGs annotated to ABCG1, DHCR24, and MYLIP were common to ischemic heart disease and type 2 diabetes (Fig 2B). We also examined where the 69 associations of interest were located in relation to CpG islands. CpG islands are clusters of methylation sites that typically occur at or near transcription start sites. Only 1 CpG was annotated to a CpG island (cg00994936), 20 were located in shores (0 to 2 kb from islands), 11 were in shelves (2 to 4 kb from islands), and the remaining 37 were annotated to the “open sea” (isolated sites outside of islands) (S8 Table).

Genetic colocalisation analyses provided weak evidence for a shared causal variant underlying methylation at cg00857282 (MYLIP) and risk of ischemic heart disease (PP = 63%; S9 Table). There was also moderate evidence for distinct causal variants underlying 10 of the 69 prevalent associations (PP > 75%).

3.3. Epigenome-wide analyses on incident disease

Using health record linkage, we tested whether CpGs measured at baseline associated with the future onset of 19 disease states. We observed 14,237 associations between baseline CpG methylation and the incidence of 11 disease states in the basic model (p < 1.9 × 10⁻⁹; Fig 3A, S10 Table). Of these, 11,305 (79.4%) and 2,657 (18.7%) were attributed to COPD and type 2 diabetes, respectively. Well-established smoking-associated probes (e.g., cg14391737 within PRSS23 and cg05575921 within AHRR) associated with the incidence of COPD, lung cancer, ischemic heart disease, stroke, pain, and/or CKD. Genomic inflation factors ranged from 0.8 to 3.8 across all basic incidence models (S11 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Epigenome-wide association studies on 19 incident disease states in Generation Scotland.

Diseases that were identified in only the basic model or only the fully adjusted model are shown in bold. COVID severity, liver cirrhosis, and ovarian cancer were present in both a basic and fully adjusted model, but there were no overlapping CpGs for these disease states in both models. (B). Ideogram showing 64 associations that were common to the basic and fully adjusted models. Full information is available in S12 Table. Image created using Biorender.com. COPD, chronic obstructive pulmonary disease; CpG, cytosine-phosphate-guanine dinucleotide; WBC, white blood cells.

https://doi.org/10.1371/journal.pmed.1004247.g003

There were 79 unique associations in the fully adjusted model, which were spread across 5 disease states (Fig 3B, S12 Table). However, only 64 associations for COPD (n = 6) and type 2 diabetes (n = 58) were present across both basic and fully adjusted models. One site was annotated to a CpG island (cg14334350), 10 were in shores, 12 were in shelves, and 41 were located in the “open sea.” Genomic inflation factors ranged from 0.8 to 1.8 across all fully adjusted incidence models and were 1.1 and 1.8 for COPD and type 2 diabetes, respectively (S11 Table). Genes annotated to CpGs that associated with COPD included ALPG, C11orf91, CPOX, GPR15, HLA-DRB5, and PRSS23. Genes annotated to CpGs that were associated with type 2 diabetes included ABCA1, ABCG1, CPT1A, SREBF1, SLC7A11, SLC7A5, and TXNIP among others (see S12 Table for full details). Only type 2 diabetes had CpGs common to cross-sectional and longitudinal analyses and reflected 17 CpGs annotated to 11 unique genes.

There was only moderate evidence for distinct causal variants underlying 11/64 incident associations (PP > 75%). No associations showed strong evidence of colocalisation (S13 Table).

As a further analysis, we examined the contribution of each of the 5 common lifestyle risk factors in attenuating the 1,340 prevalent associations and 14,237 incident associations that were brought forward to the fully adjusted stage. The findings are outlined in full in S5 Text and S14 Table. In brief, the mean attenuation in effect sizes by each of the covariates ranged from 5.5% (for body mass index) to 63.1% (for smoking). However, there was heterogeneity across disease states given their distinct risk profiles.

3.4. Pathway enrichment analysis for methylation sites associated with common disease states

The top 100 CpGs (i.e., with the smallest EWAS p-values) for each fully adjusted model were assessed for enrichment in KEGG pathways and GO terms (see Methods). Thirty-three models were considered and reflected 14 prevalent and 19 incident phenotypes (S15 Table). Significant pathways were returned only for prevalent type 2 diabetes and ischemic heart disease (FDR-corrected p-value <0.05). The overrepresented terms included cholesterol and metabolic processes as well as alcohol metabolic pathways, which may indicate residual confounding despite adjustment for self-reported alcohol consumption.

3.5. Associations between CpG methylation and disease states are robust in sensitivity analyses

Mixed-effects models that included a kinship matrix were used to account for relatedness as sensitivity analyses. Effect sizes correlated >0.99 with associations from the standard EWAS, which included related individuals (S16 and S17 Tables, S3 Fig). Further, Cox proportional hazard models are often used to conduct incidence analyses. This model relies on the proportional hazard assumption, which in effect states that the hazard ratio remains constant over time and implies that the effect of a risk variable is also constant over the length of follow-up. The assumption is supported by a nonsignificant relationship between Schoenfeld residuals and time and refuted by a significant association. Fourteen of the 64 incident associations violated the proportional hazard assumption (p < 0.05 between Schoenfeld residuals and time; S18 Table). However, we also restricted the analyses to each possible year of follow-up and found that there were minimal differences in hazard ratios between time-points that failed the assumption versus those that did not (S19 Table). This suggested the hazards were proportional over time and all associations were therefore retained. Furthermore, death was considered as a censoring event within our study rather than a competing risk. Effect sizes were correlated >0.99 when incidence models were repeated with death as a competing event, and when individuals who had died were excluded from analyses (S20 Table).

The large number of association models employed in EWAS renders it challenging to examine the potential influence of outlying values for each CpG site, particularly where multiple phenotypes are evaluated. In an effort to highlight possible influential outliers, we computed Cook’s distance measurements across all 69 prevalent associations (4 prevalent phenotypes) and 64 incident associations (2 incident phenotypes) that were present in basic and fully adjusted models. There were therefore 133 association models for which Cook’s distance was computed. Cook’s distance is a measure of the effect of deleting an observation on the estimated coefficients, and the associated plots for all 133 models are shown in S2 Appendix [34,35]. Two separate criteria were used to identify influential outliers: (i) individuals were deemed as outliers if their distance was greater than 3 times the mean distance across data points (standard rule of thumb) or (ii) a smaller subset of “extreme outliers” were identified based on visual inspection of the plots. There were between 174 to 565 outliers across models using the first criterion and 0 to 4 extreme outliers identified by the second criterion. Effect sizes were correlated 0.7 with those from the original EWAS when outliers from the first criterion were removed and 0.99 when those from the second criterion were omitted (S21 Table).

Fully adjusted models were repeated using logistic regression (prevalent disease) or Cox models (incident disease) with age and sex included as fixed-effect covariates. This differs from the main analytical strategy that used linear regression models with adjusted phenotype and methylation variables and allowed us to return effect sizes on an interpretable scale. Fig 4 shows odds ratios and hazard ratios associated with a per-1 SD increase in adjusted CpG methylation M-values for all 69 and 64 prevalent and incident disease associations (S22 Table). We also computed the Harrell’s C-statistic for each of the 64 incident associations, which is a measure of goodness of fit within survival analyses. Specifically, we calculated the difference between the C-statistic between a fully adjusted model with and without each CpG of interest. The model without the CpG included age, sex, estimated blood cell proportions, population structure, and 5 common lifestyle factors as outlined previously. The C-statistic from this model was 0.87 and 0.80 for COPD and type 2 diabetes, respectively. All CpGs increased the concordance index. The increment obtained from CpGs ranged from 0.1% to 1.2% (for cg00163198, type 2 diabetes) across all 64 loci (S22 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Blood CpGs associated with prevalent or incident disease states showing effect sizes on interpretable scale.

Effect sizes were reestimated using logistic regression (prevalent disease, blue points) or Cox proportional hazards models (incident disease, violet points) to return more interpretable effect sizes. Effect sizes represent a per-1 SD increase in age-, sex-, and experimental batch-adjusted CpG methylation M-values (or age- and batch-adjusted for breast cancer). CpGs shown were significant in both basic and fully adjusted models. Odds ratios and hazard ratios are detailed in S22 Table. CI, confidence interval; CpG, cytosine-phosphate-guanine dinucleotide; SD, standard deviation.

https://doi.org/10.1371/journal.pmed.1004247.g004

3.6. Structured literature review on existing epigenome-wide analyses of common diseases

We performed a structured review of the literature to identify blood-based EWAS on the 19 disease states considered in our study (n = 54 studies; Fig 5). Characteristics for each of the 54 studies are outlined (S23 Table). The studies were deemed to be of high quality. However, there was a high risk of selection bias among epigenome-wide analyses as well as attrition bias (i.e., in the incidence analyses). Fourteen disease states had at least 1 EWAS reported in the literature. The number of studies ranged from 1 (for long COVID) to 7 (for type 2 diabetes and lung cancer) (S24 and S25 Tables). Four studies used the Illumina 27k array (7.4%), 36 used the 450k array (66.7%), 12 employed the EWAS array (22.2%), and 2 implemented alternative arrays (Infinium Multi-Ethnic Global-8 and PyroMark Q24, 3.7%). Sixteen studies examined incident disease, while the remaining 38 focused on prevalent disease.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Look-up and replication analyses within EWAS on common disease states.

A structured literature search was performed to identify existing EWAS on 19 common disease states (either prevalent or incident). (1) We first determined whether associations in our study replicated those of previous studies. We focussed only on associations that were common to basic and fully adjusted models. There were 69 prevalent associations across 4 conditions (breast cancer, CKD, ischemic heart disease, and type 2 diabetes), and 64 incident associations across 2 conditions (COPD and type 2 diabetes). We found that 11/69 prevalent associations and 8/64 incident associations were reported in the literature. (2) We then turned our attention to the existing studies and asked whether studies that examined the same trait (e.g., incident type 2 diabetes) reported the same loci in their studies. We omit our study here as we are only interested in the previous literature. We required that a CpG site was reported in at least 2 studies that examined the same trait. There was a limited amount of replication in the literature as indicated in the right-hand side of the figure. Image created using Biorender.com. CKD, chronic kidney disease; CpG, cytosine-phosphate-guanine dinucleotide; COPD, chronic obstructive pulmonary disease; IBD, inflammatory bowel disease.

https://doi.org/10.1371/journal.pmed.1004247.g005

First, we performed look-up analyses to determine whether CpGs identified in our study were previously reported at significance thresholds deemed by each individual study. Only 11/69 prevalent associations in this study (including 1 for CKD and 10 for type 2 diabetes) and 8/64 incident associations (for type 2 diabetes only) were reported in the literature (at p < 2 × 10⁻⁵, which represented the least conservative threshold across studies for these traits; Fig 5). The replicated associations for type 2 diabetes implicated genes including ABCG1, CPT1A, SREBF1, and TXNIP.

Second, we assessed how well previous studies that examined the same trait (e.g., the 7 studies on type 2 diabetes) agreed with one another in terms of locus discovery. The present study was not included in this analysis as here we were interested only in the previous literature. A CpG was considered to be replicated in the literature if 2 or more studies reported it as significant at the threshold defined in their study. As different arrays may not have the same CpG sites, we also considered whether a given gene was replicated in at least 2 studies examining the same condition. There were 10 disease states that were available for testing (i.e., had 2 or more studies with available summary statistic data). The number of unique CpGs that were reported as significant by the authors ranged from 7 (for COPD) to 2,746 (for ovarian cancer). Six of the 10 disease states had evidence of replication across existing studies with respect to the CpGs identified by EWAS. They were IBD (1.1% of CpGs replicated), stroke (1.8%), ovarian cancer (2.2%), CKD (5.2%), type 2 diabetes (6.5%), and lung cancer (16.8%) (Fig 5). Similar percentages were observed for genes, with the exception of CKD, which had no common genes across studies as all of the replicated CpGs were intergenic (S25 Table).