Pulmonary Function and Blood DNA Methylation: A Multiancestry Epigenome-Wide Association Meta-analysis

Abstract

Rationale

Methylation integrates factors present at birth and modifiable across the lifespan that can influence pulmonary function. Studies are limited in scope and replication.

Objectives

To conduct large-scale epigenome-wide meta-analyses of blood DNA methylation and pulmonary function.

Methods

Twelve cohorts analyzed associations of methylation at cytosine-phosphate-guanine probes (CpGs), using Illumina 450K or EPIC/850K arrays, with FEV₁, FVC, and FEV₁/FVC. We performed multiancestry epigenome-wide meta-analyses (total of 17,503 individuals; 14,761 European, 2,549 African, and 193 Hispanic/Latino ancestries) and interpreted results using integrative epigenomics.

Measurements and Main Results

We identified 1,267 CpGs (1,042 genes) differentially methylated (false discovery rate, <0.025) in relation to FEV₁, FVC, or FEV₁/FVC, including 1,240 novel and 73 also related to chronic obstructive pulmonary disease (1,787 cases). We found 294 CpGs unique to European or African ancestry and 395 CpGs unique to never or ever smokers. The majority of significant CpGs correlated with nearby gene expression in blood. Findings were enriched in key regulatory elements for gene function, including accessible chromatin elements, in both blood and lung. Sixty-nine implicated genes are targets of investigational or approved drugs. One example novel gene highlighted by integrative epigenomic and druggable target analysis is TNFRSF4. Mendelian randomization and colocalization analyses suggest that epigenome-wide association study signals capture causal regulatory genomic loci.

Conclusions

We identified numerous novel loci differentially methylated in relation to pulmonary function; few were detected in large genome-wide association studies. Integrative analyses highlight functional relevance and potential therapeutic targets. This comprehensive discovery of potentially modifiable, novel lung function loci expands knowledge gained from genetic studies, providing insights into lung pathogenesis.

At a Glance Commentary

Scientific Knowledge on the Subject

DNA methylation can influence pulmonary function. Data on blood DNA methylation and pulmonary function are relatively few with minimal replication.

What This Study Adds to the Field

This large-scale multiancestry study of epigenome-wide DNA methylation and pulmonary function identified many novel loci, mostly not discovered by genetic studies. Various integrative analyses enhance the functional and clinical relevance of our findings, including potential therapeutic targets.

Pulmonary function traits, including FEV₁, FVC, and their ratio (FEV₁/FVC), assess the physiologic state of the lungs and provide the basis for diagnosing chronic obstructive pulmonary disease (COPD). They predict morbidity and mortality in the general population after accounting for other risk factors, even within the normal range (1, 2). The mechanisms for these associations remain largely unknown.

Adult pulmonary function reflects environment and genetics. Various exposures, most notably cigarette smoking, reduce lung function. Large-scale genome-wide association studies (GWASs) have implicated more than 300 loci (3, 4); much of the variability remains unexplained. Epigenetic DNA modifications reflect genetics and exposures over the life course and can identify genes influencing pulmonary function. Methylation is the most studied epigenetic modification due to high-throughput, reproducible platforms with reasonable genome-wide coverage. Several epigenome-wide association studies (EWASs) using Illumina 27K (5), 450K (6–8), and EPIC/850K (9) platforms have identified pulmonary function–related cytosine-phosphate-guanine probes (CpGs). However, replication has been limited. Most were studies in European ancestry populations; no large-scale multiancestry study has been published.

We performed a meta-analysis of coordinated EWAS results from 16 separate analyses from 12 cohorts (17,503 individuals, including 14,761 European, 2,549 African, and 193 Hispanic/Latino ancestries) to identify CpGs differentially methylated in relation to pulmonary function. To provide insight into functional impacts, we evaluated associations between identified CpGs and nearby gene expression in paired blood DNA methylation and total blood RNA transcriptome data. Using integrative epigenomic methods, we assessed enrichment of regulatory elements in our blood findings across tissue types, including lung. Mendelian randomization (MR) and colocalization analyses were used to provide further functional interpretation. For clinical implications, we explored whether implicated genes were targets of drugs approved or under investigation. This large-scale multiancestry study of blood DNA methylation and pulmonary function identified numerous loci not found in GWASs, increasing our understanding of mechanisms regulating pulmonary function. Some preliminary results of this study were previously reported in the form of an abstract (10).

Methods

Further details regarding the methods are provided in the online supplement. The population included 17,503 adults (⩾40 yr old) from ALHS (Agricultural Lung Health Study), ARIC (Atherosclerosis Risk in Communities Study), CHS (Cardiovascular Health Study), FHS (Framingham Heart Study), GS (Generation Scotland), LifeLines, LBC (Lothian Birth Cohort), MESA (Multi-Ethnic Study of Atherosclerosis), RS (Rotterdam Study), and TwinsUK within the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) Consortium framework. Trained staff in each study measured prebronchodilator pulmonary function (FEV₁ and FVC in ml) (11). Methylation was assessed in blood drawn at the same visit at spirometry, except for CHS, where the difference was 1 year. Ten studies used the Illumina HumanMethylation450 (∼480,000 CpGs), and six used the newer MethylationEPIC (∼850,000 CpGs). Each cohort retained autosomal CpGs after preprocessing and filtering.

Association Analyses

Each study assessed effects of methylation on spirometric traits after adjusting for age, age squared, sex, height, height squared, weight (for FVC only), smoking (never, former, or current), pack-years, and estimated cell type proportions (12) using robust linear regression to help account for potential heteroskedasticity and influential outliers. In addition, studies adjusted for analytic batch, ancestry principal components (calculated from genome-wide genotypes), study site, and selection factor or accounted for family structure when appropriate. Cohorts with more than one ancestry group performed analyses separately by ancestry. We combined study-specific results using inverse variance–weighted fixed-effects models (13, 14). We conducted two separate EWAS meta-analyses (450K and EPIC-unique), setting a genome-wide significance threshold of false discovery rate (FDR) <0.025 (0.05/2 meta-analyses) (15). Unless otherwise noted, genome-wide significant CpGs refer to those with FDR <0.025 from either meta-analysis. We examined differentially methylated regions using DMRcate (16).

CpG Annotation and Filtering

We used Illumina (17, 18), Zhou and colleagues (19), and Homer version 4.9.1 (20) (genome build GRCh37/hg19) for CpG annotation. From CpGs meeting genome-wide significance in meta-analyses, we removed those previously reported as potentially problematic (19). Using leave-one-out meta-analyses, we identified and removed CpGs with associations driven by a single study. Specifically, we did not consider a CpG significant genome-wide when the association did not meet at least nominal significance (P  <  0.05) in meta-analysis, with consistent direction, after leaving out one study. Remaining CpGs were subjected to downstream analyses.

Functional Downstream Analyses

We tested for enrichment of genomic features (CpG islands, shores, shelves, promoters, and transcription factor [TF] binding sites). Using the eFORGE integrative epigenomics approach (21–23), we explored whether our lung function–associated CpGs were enriched in regulatory elements from the Roadmap Epigenomics Consortium (24) across more than 20 tissue types. To gain further biological insights, we conducted pathway analyses (25, 26).

Cis-expression quantitative trait methylation analysis

We assessed whether our significant CpGs associate with transcription of nearby genes using paired whole-blood 450K methylation and transcriptome data from FHS (27) and the BIOS (Biobank-based Integrative Omics Study) Consortium (28).

Methylation quantitative trait loci, MR, and colocalization

To examine whether our significant CpGs were methylation quantitative trait loci (mQTLs), we used the Genetics of DNA Methylation Consortium (GoDMC) database (29). To investigate causality, we performed MR (30) and colocalization (31) analyses.

Druggable targets

To explore clinical relevance, we searched for approved or experimental drugs targeting genes implicated in our meta-analyses using the ChEMBL database (32).

COPD

We examined whether CpGs significantly related to pulmonary function were associated in our data with COPD, defined using prebronchodilator spirometry: FEV₁ <80% predicted (33) and FEV₁/FVC <0.7. Noncases had FEV₁ ⩾80% predicted and FEV₁/FVC ⩾0.7 (34, 35).

Replication and Validation

To replicate our findings, we looked up our significant CpGs in recent EWASs of lung function (6, 8). To validate previous findings (5–9), we examined CpGs reported as related to pulmonary function, after multiple-testing correction and smoking adjustment, in our meta-analyses. We looked up CpGs related to COPD in lung tissue (36) in our lung function meta-analyses.

Additional Analyses

To assess whether methylation findings were driven by genetic variants for lung function, we examined our significant findings after additional adjustment for polygenic risk scores (4, 37) in ALHS. To address possible residual confounding by smoking, we evaluated whether our significant CpGs overlapped with CpGs related to current smoking (38). We also adjusted for cg05575921 (AHRR), a biomarker of lifetime smoking (39), and two additional biomarkers: cg13039251 (PDZD2) and cg03636183 (F2RL3) (40, 41). We conducted meta-analyses separately for never smokers (n = 8,830) and ever smokers (n = 8,673) to evaluate whether pulmonary function–related methylation differs by smoking. In separate meta-analyses by European (n = 14,761) and African ancestry (n = 2,549), we considered consistency across ancestries and explored ancestry-specific signals.

Results

We performed a meta-analysis of data from 17,503 participants (16 separate analyses from 12 cohorts; 12 European ancestry, 3 African ancestry, and 1 Hispanic/Latino ancestry) (Table 1; see study-level characteristics in Tables E1–E3 in the online supplement). We included up to 865,971 CpGs analyzed in at least three studies: 473,215 (93% also on EPIC) in the 450K EWAS meta-analysis and 392,756 in the EPIC-unique meta-analysis (workflow in Figure 1).

Table 1.

Characteristics of Participating Studies (N = 17,503 Participants)

Study	Country	Methylation array	Number of Participants by Ancestry
			European	African	Hispanic/Latino
ALHS	United States	EPIC	2,268
ARIC	United States	450K	787	2,261
CHS	United States	450K	218	181
FHS	United States	450K	3,205
GS Set 1	United Kingdom	EPIC	1,700
GS Set 2	United Kingdom	EPIC	2,954
LBC1921	United Kingdom	450K	435
LBC1936	United Kingdom	450K	905
LifeLines	The Netherlands	450K	1,155
MESA	United States	EPIC	246	107	193
RS	The Netherlands	450K	716
TwinsUK	United Kingdom	450K	172
Total			14,761	2,549	193

Definition of abbreviations: ALHS = Agricultural Lung Health Study; ARIC = Atherosclerosis Risk in Communities Study; CHS = Cardiovascular Health Study; FHS = Framingham Heart Study; GS = Generation Scotland; LBC = Lothian Birth Cohort; MESA = Multi-Ethnic Study of Atherosclerosis; RS = Rotterdam Study.

Figure 1.

Overview of our epigenome-wide association study (EWAS) meta-analyses on pulmonary function. (A) Each study examined associations between DNA methylation and pulmonary function. Participating studies were ALHS (Agricultural Lung Health Study), ARIC (Atherosclerosis Risk in Communities), CHS (Cardiovascular Health Study), FHS (Framingham Heart Study), GS (Generation Scotland), LifeLines, LBC (Lothian Birth Cohort), MESA (Multi-Ethnic Study of Atherosclerosis), RS (Rotterdam Study), and TwinsUK. Our EWAS meta-analyses included datasets from three ancestries: European ancestry (EA), African ancestry (AA), and Hispanic/Latino ancestry (H/L). (B) Two separate meta-analyses were conducted: 450K EWAS meta-analysis (17,503 individuals) and EPIC-unique EWAS meta-analysis (7,468 individuals). (C) Functional follow-up analyses included cis-expression quantitative trait methylation (eQTM) analyses using paired blood methylation and transcriptome data from FHS and the BIOS (Biobank-based Integrative Omics Study) Consortium, eFORGE DNase I hypersensitive site (DHS) analysis using Roadmap Epigenomics data, pathway analyses using gene sets from Kyoto Encyclopedia of Genes and Genomes, druggable targets analysis, and Mendelian randomization (MR) and colocalization analyses. (D) We replicated 12% of our significant findings in an EWAS of pulmonary function (N ≅ 2,000 EA participants) (false discovery rate [FDR], <0.05); an additional 49% were significant at P < 0.05. Previously reported cytosine-phosphate-guanine probes (CpGs) were validated: 113 (49%) of 231 CpG–trait pairs from five cross-sectional studies and 29 (71%) of 41 CpG–trait pairs from the two longitudinal studies.

Pulmonary Function–related CpGs

We identified 1,267 CpGs (1,042 genes) significantly differentially methylated (FDR, <0.025) in relation to pulmonary function (Table E4), including 164 from EPIC-unique meta-analyses and 85% associated with only one trait (Table 2, Figure E1). Of the 1,042 implicated genes, 24% contained multiple genome-wide significant CpGs. Tables 3–5 display the top 30 CpGs for each trait. Of 1,451 genome-wide significant associations (FDR, <0.025) with any of the three traits (Tables E5–E7), after removing 70 driven by a single study, 165 met Bonferroni correction (Table E8). We provide graphic representation of EWAS meta-analysis results: Miami (Figures 2 and E2) and QQ plots (λ values 1–1.3, supporting minimal inflation; Figure E3). For significant CpGs, we plotted leave-one-out meta-analysis results (Figure E4), study-specific results (forest plots in Figure E5), and distributions (Figure E6). Using DMRcate (16), we identified 2,806 differentially methylated regions associated (FDR, <0.01) with FEV₁, FVC, or FEV₁/FVC (Table E9); ∼25% contained a genome-wide significant CpGs.

Table 2.

Number of Cytosine-Phosphate-Guanine Probes Differentially Methylated (False Discovery Rate <0.025) in Relation to FEV₁, FVC, and FEV₁/FVC in Multiancestry Meta-Analyses

Trait	450K and EPIC-Unique Meta-Analyses (N  =  17,503)	450K Meta-Analyses (N  =  17,503)	EPIC-Unique Meta-Analyses (n  =  7,468)
FEV1 only	935	783	152
FVC only	47	46	1
FEV1/FVC only	101	98	3
Both FEV1 and FVC	168	160	8
Both FEV1 and FEV1/FVC	15	15	0
Both FVC and FEV1/FVC	1	1	0
Total	1,267	1,103	164

Table 3.

Top 30 Cytosine-Phosphate-Guanine Probes Differentially Methylated (False Discovery Rate <0.025) in Relation to FEV₁ in Multiancestry Meta-Analysis of Content on the 450K Array (16 Separate Analyses from 12 Cohorts; 17,503 Individuals) or Unique to EPIC (6 Studies; 7,468 Individuals), Sorted by Chromosomal Position

Chromosomal Position*	CpG Probe	Regression Coefficient	SE	P Value	Mean Methylation†	Gene Name‡	Meta-Analysis
1:2161049	cg05603985	7.828	1.168	2.1 × 10−11	0.252	SKI	450K
1:55353706	cg17901584	4.256	0.614	4.1 × 10−12	0.469	DHCR24	450K
1:109757585	cg03725309	7.625	1.062	7.0 × 10−13	0.138	SARS	450K
1:120255992	cg14476101	3.533	0.538	5.2 × 10−11	0.608	PHGDH	450K
1:145441552	cg19693031	4.441	0.661	1.8 × 10−11	0.729	TXNIP	450K
2:65225988	cg23831876	7.772	1.169	3.0 × 10−11	0.833	SLC1A4	450K
2:233284661	cg21566642	3.786	0.552	6.8 × 10−12	0.500	ECEL1P1 §	450K
3:101901234	cg12992827	5.112	0.793	1.2 × 10−10	0.702	ZPLD1 §	450K
3:185538892	cg24960291	4.853	0.727	2.4 × 10−11	0.597	IGF2BP2	450K
4:57947735	cg15696506	4.127	0.610	1.4 × 10−11	0.505	IGFBP7	450K
4:139162808	cg06690548	6.039	0.694	3.3 × 10−18	0.849	SLC7A11	450K
5:373378	cg05575921	5.051	0.581	3.6 × 10−18	0.815	AHRR	450K
5:159428643	cg18394552	−4.348	0.657	3.7 × 10−11	0.597	TTC1 ‖	450K
6:166970252	cg17501210	4.626	0.687	1.7 × 10−11	0.704	RPS6KA2	450K
8:121597619	cg01198738	6.235	0.908	6.5 × 10−12	0.478	SNTB1	EPIC-unique
11:68607622	cg00574958	16.308	2.023	7.6 × 10−16	0.069	CPT1A	450K
11:68607737	cg17058475	9.571	1.307	2.5 × 10−13	0.081	CPT1A	450K
11:102189303	cg19120513	9.894	1.390	1.1 × 10−12	0.516	BIRC3	EPIC-unique
12:11898284	cg07986378	4.130	0.614	1.8 × 10−11	0.605	ETV6	450K
12:104853274	cg06647068	4.751	0.677	2.3 × 10−12	0.285	CHST11	450K
13:79968324	cg16969872	5.421	0.752	5.8 × 10−13	0.703	RBM26	450K
14:74227441	cg10919522	5.497	0.822	2.3 × 10−11	0.216	C14orf43	450K
16:75079000	cg08761535	8.906	1.358	5.4 × 10−11	0.793	ZNRF1	450K
17:76354621	cg18181703	6.180	0.704	1.7 × 10−18	0.454	SOCS3	450K
17:76354934	cg11047325	4.617	0.706	6.3 × 10−11	0.570	SOCS3	EPIC-unique
19:45252955	cg26470501	6.869	0.861	1.4 × 10−15	0.504	BCL3	450K
19:47287778	cg22304262	5.543	0.829	2.2 × 10−11	0.735	SLC1A5	450K
19:47287964	cg02711608	9.412	1251	5.2 × 10−14	0.173	SLC1A5	450K
22:31686097	cg08548559	4.534	0.681	2.9 × 10−11	0.210	PIK3IP1	450K
22:50327986	cg09349128	6.998	0.931	5.7 × 10−14	0.280	CITF22-49E9.3 §	450K

Definition of abbreviation: CpG = cytosine-phosphate-guanine probe.

Top 30 CpGs based on meta-analysis P values. Individual study results were obtained from robust linear regression with methylation as the predictor and FEV₁ as the outcome. Covariates included age, sex, height, age squared, height squared, smoking status (never, former, or current), pack-years, and estimated cell type proportions. Study-specific covariates included study center, selection factor, ancestry principal components, batch variables, and family structure when appropriate. Regression coefficients represent milliliter differences in FEV₁ per 1% difference in methylation. Table E5 contains complete results (false discovery rate, <0.025).

^*Genome build GRCh37/hg19.

^† Weighted average methylation across participating studies at the specified CpG.

^‡ Gene names from the Illumina annotation (17, 18), Zhou and colleagues (19), or Homer version 4.9.1 (20).

^§ Gene names from Zhou and colleagues (19).

^‖ Gene names (within ±2 Mb) from Homer version 4.9.1 (20).

Table 5.

Top 30 Cytosine-Phosphate-Guanine Probes Differentially Methylated (False Discovery Rate <0.025) in Relation to FEV₁/FVC in Multiancestry Meta-Analysis of Content on the 450K Array (16 Separate Analyses from 12 Cohorts; 17,503 Individuals) or Unique to EPIC (6 Studies; 7,468 Individuals), Sorted by Chromosomal Position

Chromosomal Position*	CpG Probe	Regression Coefficient	SE	P Value	Mean Methylation†	Gene Name‡	Meta-Analysis
1:92947588	cg09935388	0.00045	0.00008	9.9 × 10−9	0.726	GFI1	450K
2:8343710	cg23079012	0.00116	0.00019	1.7 × 10−9	0.94	LINC00298;LINC00299 §	450K
2:70008161	cg05155595	0.00075	0.00013	1.6 × 10−8	0.641	ANXA4	450K
2:233284661	cg21566642	0.00058	0.00009	1.2 × 10−11	0.5	ECEL1P1 §	450K
3:98251294	cg19859270	0.00171	0.00027	3.5 × 10−10	0.887	GPR15	450K
4:8174148	cg09390241	−0.00059	0.00010	4.0 × 10−9	0.667	ABLIM2 ‖	450K
4:109038130	cg12623364	0.00086	0.00015	1.3 × 10−8	0.188	LEF1	450K
5:373378	cg05575921	0.00112	0.00009	9.1 × 10−35	0.815	AHRR	450K
5:393347	cg17287155	0.00120	0.00021	9.7 × 10−9	0.885	AHRR	450K
5:393366	cg04551776	0.00096	0.00017	4.0 × 10−8	0.771	AHRR	450K
5:150161299	cg14580211	0.00071	0.00012	4.0 × 10−9	0.682	C5orf62	450K
6:30720203	cg24859433	0.00139	0.00024	6.6 × 10−9	0.834	IER3 ‖	450K
6:167536184	cg05094429	−0.00063	0.00012	1.0 × 10−7	0.673	CCR6	450K
9:108005349	cg01692968	0.00052	0.00009	9.7 × 10−9	0.323	SLC44A1 ‖	450K
9:134280803	cg14264316	0.00050	0.00009	1.7 × 10−8	0.601	PRRC2B §	450K
11:44626750	cg01199327	−0.00064	0.00012	7.5 × 10−8	0.842	CD82	450K
11:86510915	cg11660018	0.00070	0.00012	3.0 × 10−9	0.518	PRSS23	450K
12:54677008	cg02583484	0.00072	0.00013	6.1 × 10−8	0.295	HNRNPA1;HNRPA1L-2	450K
14:92979577	cg26829189	0.00102	0.00017	1.3 × 10−9	0.542	RIN3	EPIC-unique
14:92981121	cg03345232	0.00059	0.00011	2.7 × 10−8	0.578	RIN3	450K
14:92981227	cg12072028	0.00033	0.00006	1.1 × 10−7	0.699	RIN3	450K
14:94547496	cg20554312	−0.00568	0.00084	1.4 × 10−11	0.021	DDX24;IFI27L1	450K
15:45028270	cg10439456	0.00043	0.00008	5.5 × 10−8	0.351	TRIM69	450K
16:8985593	cg08065963	−0.00091	0.00016	2.0 × 10−8	0.694	CARHSP1 ‖	450K
16:8985638	cg05946118	−0.00095	0.00017	2.5 × 10−8	0.708	CARHSP1 ‖	450K
17:80872461	cg10310700	−0.00074	0.00014	1.0 × 10−7	0.817	TBCD	450K
19:17000585	cg03636183	0.00076	0.00010	1.1 × 10−13	0.655	F2RL3	450K
20:5931325	cg20225569	0.00315	0.00057	4.2 × 10−8	0.018	TRMT6;MCM8	EPIC-unique
21:43656587	cg06500161	0.00088	0.00016	2.3 × 10−8	0.593	ABCG1	450K
22:30639979	cg23635663	−0.00163	0.00022	3.9 × 10−13	0.908	LIF	450K

For definition of abbreviations, see Table 3.

Top 30 CpGs based on meta-analysis P values. Individual study results were obtained from robust linear regression with methylation as the predictor and FEV₁/FVC as the outcome. Covariates included age, sex, height, age squared, height squared, smoking status (never, former, or current), pack-years, and estimated cell type proportions. Study-specific covariates included study center, selection factor, ancestry principal components, batch variables, and family structure when appropriate. Methylation values were between 0 (unmethylated) and 1 (methylated). Regression coefficients represent differences in FEV₁/FVC ratio per 1% difference in methylation. Table E7 contains complete results (false discovery rate, <0.025).

^*Genome build GRCh37/hg19.

^† Weighted average methylation across participating studies.

^‡ Gene names from the Illumina annotation (17, 18), Zhou and colleagues (19), or Homer version 4.9.1 (20).

^§ Gene names from Zhou and colleagues (19).

^‖ Gene names (within ±2 Mb) from Homer version 4.9.1 (20).

Figure 2.

Visualization of 450K epigenome-wide association study meta-analysis results (N = 17,503 participants). (A) Miami plot for FEV₁, with each dot representing the −log₁₀(P value) of a single cytosine-phosphate-guanine probe (CpG). Each plot has two panels: upper (P.Pos) for association results with positive regression coefficients and lower (P.Neg) for association results with negative regression coefficients, with −log₁₀(P value) on the y-axis and 22 chromosomes on the x-axis. Horizontal lines depict P value cutoffs for statistical significance after multiple-testing correction: Bonferroni and Benjamini-Hochberg false discovery rate (FDR). CpGs having uncorrected P > 0.05 were not displayed. (B) Same for FVC. (C) Same for FEV₁/FVC. JAK-STAT = Janus kinase/signal transducer and activator of transcription; MAPK = mitogen-activated protein kinase.

Table 4.

Top 30 Cytosine-Phosphate-Guanine Probes Differentially Methylated (False Discovery Rate <0.025) in Relation to FVC in Multiancestry Meta-Analysis of Content on the 450K Array (16 Separate Analyses from 12 Cohorts; 17,503 Individuals) or Unique to EPIC (6 Studies; 7,468 Individuals), Sorted by Chromosomal Position

Chromosomal Position*	CpG Probe	Regression Coefficient	SE	P Value	Mean Methylation†	Gene Name‡	Meta-Analysis
1:120255992	cg14476101	4.021	0.625	1.3 × 10−10	0.608	PHGDH	450K
1:145441552	cg19693031	5.244	0.784	2.3 × 10−11	0.729	TXNIP	450K
4:139162808	cg06690548	6.804	0.838	4.6 × 10−16	0.849	SLC7A11	450K
6:36326677	cg03149958	5.440	0.953	1.1 × 10−8	0.783	ETV7 §	450K
6:166970252	cg17501210	5.355	0.805	2.9 × 10−11	0.704	RPS6KA2	450K
7:71800412	cg00277397	5.943	1.056	1.8 × 10−8	0.739	CALN1	450K
8:103937374	cg19589396	5.074	0.854	2.8 × 10−9	0.688	KB-1507C5.2;RPL5P24 §	450K
8:121597619	cg01198738	6.141	1.055	5.9 × 10−9	0.478	SNTB1	EPIC-unique
8:134066590	cg17088014	5.510	0.978	1.8 × 10−8	0.348	SLA;TG	450K
9:111885602	cg13661827	4.390	0.781	1.9 × 10−8	0.429	TMEM245 ‖	450K
11:68607622	cg00574958	19.982	2.483	8.5 × 10−16	0.069	CPT1A	450K
11:68607737	cg17058475	10.018	1.573	1.9 × 10−10	0.081	CPT1A	450K
11:102189303	cg19120513	10.088	1.587	2.1 × 10−10	0.516	BIRC3	EPIC-unique
12:104853274	cg06647068	4.343	0.773	2.0 × 10−8	0.285	CHST11	450K
13:79968324	cg16969872	5.571	0.875	1.9 × 10−10	0.703	RBM26	450K
15:40620444	cg04847110	8.201	1.440	1.2 × 10−8	0.789	INAFM2 ‖	450K
15:59587546	cg24263283	6.125	1.053	6.0 × 10−9	0.815	MYO1E	450K
15:64290807	cg07037944	8.433	1.502	2.0 × 10−8	0.199	DAPK2	450K
15:91455407	cg11183227	−5.691	0.996	1.1 × 10−8	0.808	MAN2A2	450K
16:3030649	cg02386244	−35.453	6.066	5.1 × 10−9	0.019	PKMYT1	450K
16:30410051	cg00711896	−7.818	1.338	5.2 × 10−9	0.89	ZNF48	450K
16:75079000	cg08761535	9.763	1.613	1.4 × 10−9	0.793	ZNRF1	450K
17:27333185	cg04614997	−17.937	3.132	1.0 × 10−8	0.03	SEZ6	450K
17:76354621	cg18181703	6.510	0.821	2.1 × 10−15	0.454	SOCS3	450K
17:76354934	cg11047325	4.605	0.817	1.7 × 10−8	0.57	SOCS3	EPIC-unique
18:47901430	cg16196758	−10.705	1.831	5.0 × 10−9	0.016	SKA1	450K
19:1423902	cg00994936	−7.423	1.249	2.8 × 10−9	0.831	DAZAP1	450K
19:45252955	cg26470501	6.801	0.998	9.3 × 10−12	0.504	BCL3	450K
19:47287964	cg02711608	9.360	1.455	1.2 × 10−10	0.173	SLC1A5	450K
22:50327986	cg09349128	6.838	1.072	1.8 × 10−10	0.280	CITF22-49E9.3 §	450K

For definition of abbreviations, see Table 3.

Top 30 CpGs based on meta-analysis P values. Individual study results were obtained from robust linear regression with methylation as the predictor and FVC as the outcome. Covariates included age, sex, height, age squared, height squared, weight, smoking status (never, former, or current), pack-years, and estimated cell type proportions. Study-specific covariates included study center, selection factor, ancestry principal components, batch variables, and family structure when appropriate. Regression coefficients represent milliliter differences in FVC per 1% difference in methylation. Table E6 contains complete results (false discovery rate, <0.025).

^*Genome build GRCh37/hg19.

^† Weighted average methylation across participating studies.

^‡ Gene names from the Illumina annotation (17, 18), Zhou and colleagues (19), or Homer version 4.9.1 (20).

^§ Gene names from Zhou and colleagues (19).

^‖ Gene names (within ±2 Mb) from Homer version 4.9.1 (20).

Functional Impacts

Notably, in our significant findings, TF binding sites and promoter regions (for FEV₁, the trait with the largest number of findings) were enriched (Table E10), supporting potential impacts on transcription. Integrative epigenomic analyses (eFORGE) for FEV₁ highlighted enrichment of DNase I hotspots in blood and lung (Figure 3A). Enrichment in blood and fetal lung were distinct signals (Figures 3B and 3C). FEV₁-associated genetic variants also show enrichment for lung DNase I hotspots (Figure 3D) for a different set of loci when compared with FEV₁-associated CpGs (42). Because DNase I hotspots represent broad regions of accessible chromatin containing various regulatory elements, these analyses highlight functional implications.

Figure 3.

Integrative epigenomic analysis indicates potential effects on lung and blood and comparison with pulmonary function genome-wide association study (GWAS) loci. (A) FEV₁-related cytosine-phosphate-guanine probes (CpGs) (false discovery rate, <0.025): eFORGE analysis to quantify enrichment in DNase I hotspots. The x-axis represents tissue- and cell-type samples used in the analysis; the y-axis indicates enrichment (−log₁₀ P value). (B) eFORGE DNase I hotspot analysis limited to FEV₁-related CpGs in the top blood component. (C) eFORGE DNase I hotspot analysis limited to FEV₁-related CpGs in the top fetal lung component. (D) FORGE2 DNase I hotspot analysis of FEV₁-related genetic variants from the GWAS catalog.

Pathway Enrichment

Pathways relevant to pulmonary function were enriched (FDR, <0.05) (Figure 4). Several pathways overlapped across the traits, including Wnt signaling, a key developmental pathway involved in lung pathogenesis, and inflammatory pathways such as cytokine–cytokine receptor interaction.

Figure 4.

Heatmap of enriched pathways (false discovery rate [FDR], <0.05) for FEV₁, FVC, and FEV₁/FVC using the methylGSA R package. Significantly enriched (FDR, <0.05) pathways for at least one of the three traits are shown. The color spectrum is based on the P values corrected for multiple testing using Benjamini-Hochberg FDR. Darker shading indicates higher level of statistical significance. The heatmap was created in R version 3.6.1, using ComplexHeatmap package version 2.7.10. JAK-STAT = Janus kinase/signal transducer and activator of transcription; MAPK = mitogen-activated protein kinase.

Correlation with Expression

Linking our significant CpGs on the 450K to paired blood gene expression and 450K methylation data (27, 28), 97% (of 1,103 CpGs available) had at least one transcript within ±250 kb. At FDR <0.05, 56% were related to gene expression (Table E11), and at P  <  0.05, 75% were related to gene expression (P_enrichment < 2.2 × 10⁻¹⁶), supporting functional impacts on gene regulation.

mQTL, colocalization, and MR analyses

Using GoDMC (29), 798 of our 1,103 significant CpGs had at least one mQTL. The mQTLs for our EWAS findings were associated with lung function more than expected by chance: mQTLs for the FEV₁-related CpGs associated more strongly with FEV₁ than other mQTLs (P_enrichment = 7.1 × 10⁻⁷; controlling for linkage disequilibrium, minor allele frequency, and removing MHC region) (Table E12). Observed enrichment could reflect EWAS findings capturing relevant regulatory regions for the traits or potentially causal effects of the CpGs. Using MR to investigate causality, we found 78 significant associations (FDR, <0.05) (Figure E7, Table E13). Because lung function is polygenic, these associations might reflect distinct causal variants for the outcomes in linkage disequilibrium with the mQTLs. When we tested this using genetic colocalization at the mQTLs exhibiting a strong methylation–outcome association, 46 of 78 MR associations showed evidence for colocalization (posterior probability >0.8 for the hypothesis of one shared genetic variant for methylation and trait) (Figure E8, Table E13).

Druggable targets

Among the 1,042 implicated genes, 261 had bioactive molecules with druglike properties in ChEMBL, including 69 with at least one approved or candidate drug for respiratory or other conditions. Sixty-one genes had not been identified by GWAS of pulmonary function (Table E14).

COPD associations

Of the 1,267 CpGs differentially methylated in relation to pulmonary function, 73 associated with COPD in our data (1,787 cases and 11,824 noncases) at FDR <0.05, and 323 met nominal significance (P_enrichment < 2.2 × 10⁻¹⁶) (Table E15). Directions matched expectation: CpGs positively associated with pulmonary function were negatively associated with COPD and vice versa.

Replication

In a published EWAS of lung function in ∼2,000 European ancestry participants (8), 12% of our significant CpGs were associated with any of the three traits (FDR, <0.05), and an additional 49% were nominally significant (P_enrichment for both < 2.2 × 10⁻¹⁶) (Table E16). In a small Korean study (6), 14% were related to any of the three traits (P < 0.05; P_enrichment < 2.2 × 10⁻¹⁶) (Table E17).

Validating Published CpGs

Five cross-sectional studies (5–9) reported 134 CpGs associated with at least one spirometric trait at genome-wide significance. Of 231 CpG–trait associations available, 36% met FDR <0.05 and had directions of association consistent with previous reports (Table E18). At nominal P < 0.05, an additional 59 CpGs were associated with any traits. Overall, we validated 70% of previously reported CpGs in our results (P_enrichment < 2.2 × 10⁻¹⁶). For the remaining CpGs, there was another CpG annotated to the same gene that showed associations in our data (P <0.05) (Table E19). In addition, of 535 CpGs previously related to COPD in lung (36), 28% were associated with pulmonary function at nominal significance (P_enrichment < 2.2 × 10⁻¹⁶) (Table E20). Given the sparse EWAS literature on longitudinal decline in pulmonary function (7, 8), we considered findings from one study that did not adjust for smoking (8). Of 31 CpGs (7, 8) available, 24 were nominally associated with lung function in our cross-sectional data (P_enrichment < 2.2 × 10⁻¹⁶) (Table E21), including 10 showing genome-wide significance.

Adjustment for Polygenic Risk Scores

To evaluate whether our methylation findings reflect genetic susceptibility, in ALHS, we also adjusted for a polygenic risk score (4, 37) specific for each trait. Each risk score was significantly related to its trait (Table E22). Of our genome-wide significant signals available in ALHS, 54% had P < 0.05, and ∼95% remained nominally significant after adjustment for the risk score (Table E23). Regression coefficients were virtually unchanged by the adjustment (Pearson correlation, 0.998), suggesting that the EWAS meta-analysis results provide information complementary to genetics of pulmonary function.

Examination of Potential Residual Confounding by Smoking

Overlap previously reported between CpGs related to pulmonary function and those related to cigarette smoking (8) could reflect residual confounding because smoking-related CpGs are strong biomarkers that better capture lifetime smoking history than questionnaire data (39, 43). Comparing our pulmonary function–related CpGs with those related to current smoking (FDR, <0.05) in a large 450K EWAS meta-analysis (38), we found substantial overlap (Tables E5–E7). In ALHS, we confirmed that the majority (61%) of these overlapping CpGs remained at least nominally significant after adjustment for lifetime smoking biomarkers: AHRR cg05575921, PDZD2 cg13039251, and F2RL3 cg03636183 (39–41) (Table E24). Adjustment attenuated the magnitude of 71% of associations.

Associations by Smoking Status

Separate meta-analyses by smoking identified 196 differential methylation signals in never smokers and 465 in ever smokers (FDR, <0.025) (Tables E25 and E26). In both smoking groups, 66% of signals were unique, defined as having genome-wide significance (FDR, <0.025) in one group but not reaching nominal significance (P > 0.05) in the other. For 92% of these unique signals, effect estimates were different between never and ever smokers at P < 0.05. Miami, QQ, and leave-one-out meta-analyses and forest plots for smoking-stratified meta-analyses are shown in Figures E9–E12. Enriched pathways in smoking-stratified analyses (Figure E13) were similar to those overall; for FEV₁/FVC, some pathways were enriched only for never smoker findings (Figure 4). We explored potential confounding by underreported smoking and secondhand smoke exposure by also adjusting the 129 associations unique to never smokers for AHRR cg0557592 in ALHS. Approximately 50% showed nominal significance (P < 0.05); all but two remained significant after the adjustment (Table E27). Effect sizes were highly correlated (Pearson correlation, 0.99; P < 2.2 × 10⁻¹⁶), indicating minimal impact of unreported smoking on findings in never smokers.

Associations across Ancestries

Ancestry-specific analyses identified 564 differently methylated CpGs in European ancestry (Table E28) and 29 differentially methylated CpGs (Table E29) in the smaller African ancestry dataset. Figures E14–E17 contain ancestry-specific Miami, QQ, and leave-one-out meta-analyses and forest plots. Effect estimates from results in European ancestry were correlated with African ancestry results (Pearson correlation, 0.59; P <  2.2 × 10⁻¹⁶); 31% of available associations reached at least nominal significance in African ancestry–specific analyses. However, only four African ancestry findings had P < 0.05 and matching direction in Europeans, suggesting the remainder to be unique to African ancestry.

Discussion

To our knowledge, this is the first large-scale multiancestry study examining associations between blood DNA methylation and pulmonary function. We identified numerous novel CpGs differentially methylated in relation to pulmonary function. We found evidence of replication in European and Asian ancestry studies (6, 8). We also validated findings from published studies that mostly had not been replicated. Our cross-sectional findings overlapped with limited published data on lung function decline. Our large-scale study enabled identification of pulmonary function–related methylation signals unique to smoking status, an important influence on pulmonary function. Although most signals were consistent across ancestries, we also identified signals potentially unique to individuals of African ancestry. Many CpGs correlated with nearby gene expression and were enriched for key regulatory elements in both immune cells and lung, providing functional biologic relevance of our findings further supported by MR and colocalization analyses. Implicated genes include targets of approved or investigational drugs, providing potential clinical implications.

GWASs have identified >300 genetic loci for pulmonary function (3, 4). Our EWAS loci are largely distinct. Only 3% of genes we implicated were reported in the largest lung function GWAS (Table E30) (4). Adjustment for polygenic risk scores (4, 37) for pulmonary function made little change to our results, supporting the independence of our findings from GWAS signals.

Loci identified in cross-sectional GWASs in adults primarily reflect maximally attained lung growth and are skewed toward developmental genes. In contrast to genetic variants, epigenetic alterations occur throughout life in response to numerous factors impacting pulmonary function, including environmental exposures such as smoking (38), diet (44), and air pollution (45), and endogenous influences, including systemic inflammation (46) and adiposity (47). Thus, differentially methylated genes identified by EWASs complement findings from genetic studies to identify targets potentially modifiable by lifestyle or new therapeutic interventions.

Whether EWAS findings are causal for lung function is a key question but difficult to answer. We employed MR to address this. MR requires genetic instruments related to both lung function and our EWAS CpGs. Given the novelty of most loci compared with GWASs, MR has limited ability to interrogate our findings. Nonetheless, MR analysis of all the methylation–trait pairs with genetic instruments available revealed 46 CpGs that share a genetic factor with lung function, consistent with a causal effect of methylation on lung function. However, caution is required because these could reflect violation of MR pleiotropy assumptions, and few of the 46 CpGs had the multiple genetic instruments required to test this. Regardless, without inferring causality, the shared genetic variation could reflect methylation and lung function being common consequences of regulatory impacts of the genetic variants, such as influencing TF binding. Analyses showing that, compared with other mQTLs, mQTLs for our EWAS CpGs are substantially enriched for association with lung function in GWAS support a functional role for the genomic regions uncovered by the EWAS. Along with the finding in ALHS that adjustment for genetic risk scores for lung function do not alter associations, the MR analyses suggest that our EWAS findings do not reflect reverse causation (i.e., lung function influencing methylation).

Druggable target analysis enhances the clinical relevance of our findings. Several drugs identified via ChEMBL have epigenetically relevant mechanisms. For example, mocetinostat and fimepinostat, both in phase II trials, have activity against several histone deacetylase enzymes (48, 49). Notably, three candidate drugs were annotated to tumor necrosis factor receptor superfamily member 4 (TNFRSF4; also known as OX40), including the monoclonal antibody telazorlimab, a TNFRSF4 antagonist currently in phase II trials for atopic eczema.

Integrative epigenomics identifying enrichment of DNase I hotspots across blood and lung highlights that our findings in blood can inform key epigenomic mechanisms in the lung. Analyses of TF binding motifs identified some expressed in immune tissue. Some of our significant CpGs reside near or within genes that play an important role in inflammation and immunity in the lung. One example is TNFRSF4, a gene targeted by several candidate drugs, expressed in both immune and lung tissue. Annotation for chromatin states shows overlap of lung function–associated CpGs cg21815220, cg16252905, and cg17084044 with an ENCODE immune cell “strong enhancer” located proximal to the transcription start site of TNFRSF4 (Figure 5), and these CpGs also overlap with an immune cell–specific DNase I hypersensitive site. In eFORGE, both cg16252905 and cg17084044 reside in DNase I hotspots (22). In addition, cg21815220 and cg17084044 localize to the same TF motif for hypermethylated in cancer 1 (HIC1), an important epigenetically regulated gene (Figure E18). This motif is in an active enhancer proximal to TNFRSF4 that contains cg16252905. Finding differential DNA methylation in a regulatory element proximal to a lung- and blood-specific gene hints at a putative causal role in lung function for TNFRSF4, a gene not previously reported in GWAS of pulmonary function that has a role in lung-associated inflammation and immunity (50). Mechanistically, TNFRSF4 expression is thought to sustain lung tissue inflammation (50), and TNFRSF4 blockade in experimental models improves respiratory function (50). Further research is needed to confirm our findings, which constitute the first reported epidemiological association between TNFRSF4 and lung function.

Figure 5.

Pulmonary function–associated differentially methylated cytosine-phosphate-guanine probes (CpGs). (A) TNFRSF4 gene browser shot displaying (in order, starting from the top): genomic coordinates, gene locations, GTEx RNA sequencing–based gene expression across different organs, H3K27ac peaks for seven ENCODE cell lines, ENCODE chromatin state segmentations and chromatin accessibility data, and coordinates for pulmonary function–associated CpGs cg21815220, cg16252905, and cg17084044. Both cg16252905 and cg17084044 are also located in DNase I hotspots in the eFORGE catalog. These data indicate that these three CpGs overlap with an immune enhancer near the promoter of TNFRSF4, a gene expressed in lung and immune tissue, and an enhancer and DNase I hypersensitive site from ENCODE, which were detected in immune cell samples. These three CpGs are also located within 1.5 kb of an H3K27ac peak and the transcription start site of HSP90AA1. This browser shot was generated using the University of California, Santa Cruz Genomics Institute genome browser (https://genome.ucsc.edu/) on human genome build hg19. (B) Forest plots for pulmonary function–associated CpGs cg21815220, cg16252905, and cg17084044 (linked to the genomic coordinates of these CpGs in A) indicating meta-analysis β-values and SEs across pulmonary function traits: FEV₁, FVC, and their ratio (FEV₁/FVC). Studies incorporated in the meta-analysis include ALHS (Agricultural Lung Health Study), ARIC (Atherosclerosis Risk in Communities) study, CHS (Cardiovascular Health Study), FHS (Framingham Heart Study), GS (Generation Scotland), LifeLines, LBC (Lothian Birth Cohort), MESA (Multi-Ethnic Study of Atherosclerosis), RS (Rotterdam Study), and TwinsUK.

Substantial overlap between our pulmonary function–related findings and those for COPD in our data show that examination of quantitative traits in population-based studies reveals relationships relevant to clinical outcomes such as COPD. Furthermore, we validated COPD-related CpGs identified in the lung (36), indicating that blood methylation can reflect signals in target tissue. Our results could be applied in future studies with large numbers of COPD cases to examine whether a lung function methylation score predicts that outcome.

Many CpGs associated with pulmonary function in this study and another (8) also associate with smoking (38). This led to speculation regarding whether they mediate effects of smoking on lung function (8). However, many of these overlapping CpGs are strong biomarkers of smoking that capture lifetime smoking better than pack-years from questionnaires (38, 43). One CpG (AHRR cg05575921) so strongly captures exposure that it is patented as a biomarker of lifetime smoking for insurance applications (39). Attempts to identify whether differential methylation at a given CpG that is a strong smoking biomarker mediates the biologic effects of smoking on lung function will likely produce false-positive evidence for mediation (43, 51). An alternative explanation for overlap of CpGs related to both smoking and lung function is residual confounding by lifetime smoking when adjusting for exposure using questionnaire data. We attempted to address residual confounding by adjusting for three CpGs, strong biomarkers of lifetime smoking, in one of our larger studies (ALHS), and found some attenuation of effect estimates after adjustment. Although some biologic mediation of the effect of smoking on lung function by methylation biomarkers of smoking, such as AHRR cg05575921, is possible, our results are consistent with some residual confounding. To address possible residual confounding by quantitative smoking history in another way, we examined associations in lifelong never smokers. Signals unique to never smokers are less likely due to uncontrolled confounding. However, understanding whether CpGs that are strong biomarkers of smoking are truly involved in the pathogenesis of smoking-related impairment in lung function will require deciphering the fundamental mechanisms whereby smoking alters methylation at specific sites.

This study has limitations. Assessing methylation in blood limits inference to other tissue types. Our data are cross-sectional. However, a high proportion of CpGs identified in previous longitudinal studies (7, 8) associate with lung function in our data, suggesting that cross-sectional meta-analyses can shed light on methylation predictors of decline. Individuals with European ancestry, mainly from the United States, the United Kingdom, or Northern Europe, compose 84% of our population, limiting detection of ancestry-specific signals. Because only one study had Hispanic/Latino participants, most analyses focused on European and African ancestry, limiting generalizability to other populations. Most cohorts had few individuals under 40, so we excluded this potentially interesting group. Using more widely available prebronchodilator spirometry to classify COPD is another limitation. However, this approach has been taken in previous large-scale genomic meta-analyses (3, 4). Like researchers in those genomic studies, we used the actual spirometric values adjusted for factors used in prediction equations. Although this does not allow effects of height and age to differ by sex, values were highly correlated with percent predicted values using Global Lung Function Initiative equations (ALHS Spearman correlation, 0.97; P < 2.2 × 10⁻¹⁶ for both FEV₁ and FVC). We did not perform analyses by sex. Sex-specific analyses would be of interest in future studies with larger sample sizes required for reliable interaction testing (52).

Our study has several key strengths. This is the largest multiancestry study of DNA methylation and pulmonary function to date, including studies of African ancestry, Hispanic/Latino ancestry, and European ancestry populations. The multiancestry design provided evidence for ancestry-specific signals. Our large sample size enabled determination of signals unique to ever and never smokers. Results from the EPIC array identified CpGs and genes unique to this more comprehensive platform. We confirmed our findings were independent of polygenic risk of reduced lung function. Correlation of our findings with gene expression, integrative epigenomics approaches identifying key regulatory elements enriched in both blood and lung, and MR and colocalization analyses support a functional role for the genomic regions we uncovered in EWAS. Analyses of druggable targets highlighted potential clinical utility.

In conclusion, our large-scale study comprehensively identified epigenome-wide differential methylation in blood related to pulmonary function. It extends the current literature by including newer DNA methylation arrays, different ancestry populations, greatly increased sample size, implementation of state-of-the-art in silico integrative epigenomics methods, MR and colocalization, and analysis of druggable targets. We identified many novel genes related to lung function. These are potentially modifiable targets for development of preventive and therapeutic strategies. In addition, the results can be leveraged for development of epigenomic risk scores for predictive biomarkers of lung disease. These findings provide new insights into the pathogenesis of lung function impairment and respiratory disease.