Genetic loci associated with chronic obstructive pulmonary disease overlap with loci for lung function and pulmonary fibrosis

Chronic obstructive pulmonary disease (COPD) is a leading cause of mortality worldwide¹. We performed a genetic association in 15,256 cases and 47,936 controls, with replication of select top results (P < 5×10⁻⁶) in 9,498 cases and 9,748 controls. In the combined meta-analysis, we identified 22 loci at genome-wide significance, including 13 new associations with COPD. Nine of these 13 loci have been associated with lung function in general population samples^2–7; however, 4 (EEFSEC, DSP, MTCL1, and SFTPD) are novel. We noted 2 loci shared with pulmonary fibrosis^8,9 (FAM13A and DSP) but with opposite risk alleles for COPD. None of our loci overlapped with genome-wide associations for asthma; however, one locus has been implicated in the joint susceptibility to asthma and obesity¹⁰. We also identified genetic correlation between COPD and asthma. Our findings highlight novel loci, demonstrate the importance of specific lung function loci to COPD, and identify potential regions of genetic overlap between COPD and other respiratory diseases.

COPD is characterized by persistent and progressive airflow limitation diagnosed by lung function testing¹. While cigarette smoking is the major risk factor, susceptibility is also influenced by genetics^11–13. We established the International COPD Genetics Consortium (ICGC) to coordinate efforts to find susceptibility loci¹⁴. We defined cases based on pre-bronchodilator evidence of moderate-to-severe airflow limitation by modified GOLD criteria¹⁵; controls had normal spirometry, and all analyses were adjusted for age, sex, and cigarette smoking (pack-years and smoking status). We performed a two-stage genome-wide association study (Figure 1). In Stage 1, we combined 26 cohorts (Supplementary Table 1 and 2) containing 63,192 individuals (15,256 COPD cases and 47,936 controls). We selected 79 loci with P < 5×10⁻⁶ and in analysis Stage 2, we tested them in the UK BiLEVE dataset (9,498 COPD cases and 9,748 controls) from the UK Biobank and performed an overall meta-analysis (Supplementary Table 3).

Figure 1.

Study design showing cohorts used in each stage of the analysis. ARIC = Atherosclerosis Risk in Communities Study, B58 = British 1958 Birth Cohort, CHS = Cardiovascular Health Study, COPACETIC = COPD Pathology: Addressing Critical gaps, Early Treatment & Diagnosis and Innovative Concepts, ECLIPSE = Evaluation of COPD Longitudinally to Identify Predictive Surrogate End-points, eQTL = Lung Expression Quantitative Trait Loci Study, FHS = Framingham Heart Study, KARE = Korean Association Resource project, MESA = Multi-Ethnic Study of Atherosclerosis, NETT-NAS = National Emphysema Treatment Trial/Normative Aging Study, RS = Rotterdam Study, SPIROMICS = Subpopulations and intermediate outcome measures in COPD study, EOCOPD = Boston Early-Onset COPD Study, ICGN = International COPD Genetics Network, TCGS = Transcontinental COPD Genetics Study, UK BiLEVE = UK Biobank Lung Exome Variant Evaluation; NHW = Non-Hispanic white, AA = African American, EA = European American. * Studies without genome-wide array genotyping (custom genotyping)

We identified 13 genome-wide significant (P < 5×10⁻⁸) associations in Stage 1. Following the Stage 2 analysis, an additional 9 loci achieved genome-wide significance in the overall meta-analysis (Table 1, Figure 2, Supplementary Figures 1 and 2). Analysis of only European ancestry (Supplementary Table 4) and only African ancestry (Supplementary Table 5 and Supplementary Figure 3) Stage 1 cohorts showed no unique association signals. Of the 22 genome-wide significant loci described in our study, 9 have been previously described as genome- (or exome-) wide significant in studies of COPD^13,16–19: HHIP, CHRNA5, HTR4, FAM13A, RIN3, TGFB2, GSTCD-NPNT, CYP2A6, and IL27-CCDC101. The remaining 13 loci have not been previously associated with COPD at genome- (or exome-) wide significance. Eight of these 13 loci: ADGRG6/GPR126, THSD4, ADAM19, TET2, CFDP1, AGER, ARMC2, and RARB have been previously described and replicated (Supplementary Table 6) in general population GWASs of two measures of lung function (FEV₁ and FEV₁/FVC) that are used in conjunction to diagnose COPD^{2,4–7,20,21}. One locus near PID1 was previously associated with FEV₁/FVC, but had not replicated in those studies^4,6. Four loci are newly being described as genome-wide significant in association with either COPD or lung function: EEFSEC, DSP, MTCL1, and SFTPD (Figure 3).

Table 1.

Overall study results showing 22 loci with genome-wide significant P values in overall meta-analysis following UK BiLEVE Stage 2 analysis.

rsID	Closest Gene	Variant Annotation	Locus	Risk Allele	Alt Allele	Risk Allele Frequency		Stage 1 Analysis		Stage 1 P Value	UK BiLEVE (Stage 2)		UK BiLEVE (Stage 2) P Value	Overall Meta-Analysis P Value
						Mean	Range	OR	95% CI		OR	95% CI
rs13141641	HHIP	intergenic	4q31.21	T	C	0.59	0.52–0.89	1.23	1.18–1.28	1.16E-24	1.21	1.16–1.27	8.15E-18	9.10E-41
rs17486278	CHRNA5	intronic	15q25.1	C	A	0.35	0.24–0.44	1.22	1.18–1.27	2.61E-24	1.13	1.08–1.18	2.35E-07	1.77E-28
rs7733088	HTR4	intronic	5q32	G	A	0.60	0.47–0.69	1.18	1.13–1.23	4.40E-14	1.18	1.13–1.23	1.78E-13	5.33E-26
rs9399401	ADGRG6	intronic	6q24.1	T	C	0.72	0.61–0.75	1.14	1.09–1.19	3.59E-10	1.17	1.12–1.23	6.18E-11	1.81E-19
rs1441358	THSD4	intronic	15q23	G	T	0.33	0.19–0.55	1.13	1.09–1.18	2.06E-10	1.12	1.07–1.17	6.87E-07	8.22E-16
rs6837671	FAM13A	intronic	4q22.1	G	A	0.41	0.36–0.58	1.16	1.11–1.20	1.02E-14	1.07	1.02–1.11	3.75E-03	7.48E-15
rs11727735	GSTCD	intronic	4q24	A	G	0.94	0.93–0.99	1.27	1.17–1.37	1.55E-08	1.25	1.14–1.36	4.93E-07	3.84E-14
rs754388	RIN3	intronic	14q32.12	C	G	0.82	0.80–0.86	1.20	1.14–1.26	7.07E-12	1.11	1.05–1.17	1.85E-04	4.96E-14
rs113897301	ADAM19	intronic	5q33.3	AT	A	0.17	0.05–0.19	1.20	1.13–1.26	4.52E-10	1.13	1.07–1.19	2.79E-05	1.58E-13
rs2047409*	TET2	intronic	4q24	A	G	0.62	0.22–0.65	1.10	1.06–1.15	1.58E-06	1.14	1.09–1.19	1.95E-08	2.46E-13
rs2955083	EEFSEC	intronic	3q21.3	A	T	0.88	0.85–0.89	1.20	1.12–1.27	2.00E-08	1.17	1.09–1.25	4.01E-06	4.16E-13
rs7186831*	CFDP1	intergenic	16q23.1	A	G	0.43	0.23–0.47	1.12	1.07–1.18	3.54E-06	1.12	1.07–1.17	6.63E-07	1.12E-11
rs10429950*	TGFB2	intergenic	1q41	T	C	0.73	0.22–0.77	1.12	1.07–1.16	1.83E-07	1.10	1.04–1.15	1.94E-04	1.66E-10
rs2070600*	AGER	coding	6p21.32	C	T	0.95	0.85–0.99	1.28	1.15–1.41	3.54E-06	1.21	1.10–1.32	2.96E-05	5.94E-10
rs2806356*	ARMC2	intronic	6q21	C	T	0.18	0.05–0.24	1.12	1.07–1.18	2.84E-06	1.12	1.06–1.18	6.88E-05	8.34E-10
rs16825267*	PID1	intergenic	2q36.3	C	G	0.93	0.87–0.94	1.24	1.15–1.34	5.22E-08	1.13	1.04–1.22	2.27E-03	1.68E-09
rs2076295	DSP	coding	6p24.3	T	G	0.55	0.44–0.58	1.11	1.07–1.15	4.95E-08	1.06	1.02–1.11	7.45E-03	3.97E-09
rs647097*	MTCL1	intronic	18p11.22	C	T	0.27	0.26–0.40	1.11	1.06–1.15	3.03E-06	1.09	1.04–1.14	4.66E-04	6.14E-09
rs1529672	RARB	intronic	3p24.2	C	A	0.83	0.68–0.86	1.16	1.11–1.22	2.37E-09	1.05	0.99–1.11	9.95E-02	2.47E-08
rs721917*	SFTPD	coding	10q22.3	G	A	0.42	0.39–0.63	1.09	1.05–1.13	2.11E-06	1.07	1.02–1.11	2.60E-03	2.49E-08
rs12459249*	CYP2A6	intergenic	19q13.2	C	T	0.66	0.62–0.70	1.13	1.07–1.18	2.89E-06	1.08	1.03–1.13	1.35E-03	3.42E-08

OR = odds ratio, CI = confidence interval.

^*Genome-wide significant in overall meta-analysis only

Figure 2.

Manhattan plot showing P values for Stage 1 analysis (small open diamonds) with overlay of overall meta-analysis P values for SNPs analyzed in UK BiLEVE Stage 2 analysis (filled circles). Gene names in gray are previously described COPD or lung function (FEV₁ or FEV₁/FVC) loci; black are novel loci discovered in this study. The Stage 1 cohorts with available genotyping data (Supplementary Figures 1a–v) and the UK BiLEVE cohort determined the sample size for each top variant. The red dashed line indicates the threshold for genome-wide significance (P value < 5×10⁻⁸).

Figure 3. a–d Regional association for novel loci.

LocusZoom plots showing regional association of variants at the four novel COPD loci. The point size is proportional to the sample size, where Stage 1 cohorts with available genotyping data (Supplementary Figures 1a–v) and the UK BiLEVE cohort determined the sample size for each top variant.

To explore the potential function and causal genes for our novel loci, in addition to using publicly available datasets and prioritization tools (Supplementary Table 7), we also examined a larger set of lung expression quantitative trait loci (eQTL) in 1038 subjects, including subjects with COPD²² (Supplementary Table 8). As eQTL are pervasive, we also attempted to determine whether our association signal co-localized²³ with an eQTL signal in lung tissue (Supplementary Table 9). We found strong evidence of co-localization (posterior probability > 0.8) for DSP, a major protein of desmosomes required for epidermal integrity²⁴, and MTCL1, important in epithelial-cell-specific microtubule stabilization^25,26, and expressed in respiratory epithelial cells²⁷. Variants in strong LD with our top MTCL1 variant rs647097 (NC_000018.9:g.8808464T>C) appear to have enhancer histone marks in fetal lung fibroblasts^28,29. In contrast, we found no evidence of a strong eQTL signal or co-localization at our other two novel loci. At 3q21, EEFSEC is a potential candidate, as it is a paralog of TUFM, a top blood and lung eQTL gene for the 16p11.2/IL27 COPD susceptibility locus¹⁹, recently part of a novel COPD-related pathway involving NLRX1^30–32. At 10q22, pulmonary surfactant-associated protein D (SFTPD) is the most likely candidate, as it is highly expressed in pneumocytes²⁷, and sftpd (−/−) mice develop pulmonary emphysema³³. SFTPD has been explored as a COPD biomarker³⁴, and while rs721917 (NC_000010.10:g.81706324A>G) is not an eQTL, polymorphisms in SFTPD, including rs721917, may lead to decreased surfactant protein D levels³⁵; though the association of SFTPD polymorphisms with COPD susceptibility have been inconsistent. Our analysis also led to some additional insights into other previously described loci. We found evidence of COPD association and eQTL statistical co-localization in lung tissue (posterior probability > 0.8) for THSD4, HHIP, AGER, CHRNA3, and RARB (Supplementary Table 9). Additional data on eQTLs (Supplementary Table 8), cohort-specific associations at each locus (Supplementary Figures 1a–v), fine mapping (Supplementary Note and Supplementary Table 10), causal gene (Supplementary Table 11 and 12), and other supportive analysis for previously described and novel loci can be found in the Supplementary Note.

We note that our top variant at DSP (rs2076295, NC_000006.11:g.7563232T>G) is also associated (P = 1.1×10⁻¹⁹) with pulmonary fibrosis⁸. Recently, a re-sequencing study³⁶ at the DSP locus identified a second fibrosis-associated variant, rs2744371 (NC_000006.11:g.7554174A>C) with P_fibrosis = 0.002 and P_COPD = 0.04. We also note overlap at the FAM13A locus (top fibrosis SNP⁸, rs2609255 [NC_000004.11:g.89811195G>T]; P_fibrosis = 2.2×10⁻¹¹, P_COPD = 1.9×10⁻⁷). We performed additional analysis to investigate genetic overlap using gwas-pw³⁷ (see Supplementary Note). We confirmed overlap at the DSP and FAM13A loci with a posterior probability of > 0.99, and additionally discovered overlap near MAPT/KANSL1 (top fibrosis SNP⁸, rs1981997 [NC_000017.10:g.44056767G>A]; P_fibrosis = 8.87×10⁻¹⁴, P_COPD = 4.5×10⁻³) with posterior probability of 0.84. While the MAPT/KANSL1 locus did not reach genome-wide significance in our study, we note its independent discovery in a genome-wide association in extremes of lung function⁷. Notably, for all four of these variants (in DSP [2], FAM13A, and MAPT), the fibrosis risk allele is protective for development of COPD. Emphysema, a key component of COPD, and pulmonary fibrosis are both smoking-related lung diseases that have both shared and distinct pathophysiology^38–40, though genetic loci with opposing effects have not been previously described. Additional investigation of these loci as a well as a more comprehensive assessment of genetic overlap of COPD and pulmonary fibrosis may lead to insight into both disorders.

Because our analysis relied on a spirometric definition of COPD alone, we did not specifically exclude other causes of airway obstruction such as asthma, which can overlap with COPD in adults⁴¹. To define COPD, we used pre-bronchodilator spirometry, which was available across all cohorts, and we included at least moderately affected cases (FEV₁ < 80% predicted). We examined the top set of genome-wide significant results in a subset of our largest cohorts with both pre- and post-bronchodilator data and densely imputed genotypes; overall, the effect sizes (mean difference = 0.001) and P values (mean log₁₀ P value difference = 0.18) were similar (Supplementary Table 13 and Supplementary Figures 4 and 5). In addition, a recent GWAS of FEV₁, FVC, and FEV₁/FVC did not find substantial differences including and excluding subjects with asthma⁷. In the 79 variants tested in Stage 2, we found no significant difference in the OR for COPD association when including and excluding individuals with asthma (Supplementary Figure 6).

We examined COPD associations of genome-wide significant asthma (and asthma-associated traits) loci from the NHGRI-EBI GWAS Catalog⁴² (Supplementary Table 14). We also compared our COPD association results to the GABRIEL asthma study⁴³ (Supplementary Tables 15). None of the genome-wide significant loci from asthma and COPD overlapped. Further, no asthma or COPD loci showed Bonferroni-adjusted (for number of look-ups) significant association with the other disease, though several loci showed nominal (P<0.05) significance. The 16p11.2 (CCDC101) locus has been described in the joint susceptibility to asthma and obesity¹⁰. COPD susceptibility is strongly related to cigarette smoking. Two of our loci (15q25 and 19q13) have been previously associated with smoking behavior^44,45, though we found no additional evidence of overlap in genome-wide significant variants described in the NHGRI-EBI GWAS Catalog⁴² and Tobacco and Genetics Consortium GWAS⁴⁵ (Supplementary Tables 16–18). We additionally evaluated overlap of our top 22 loci with COPD comorbidities (Supplementary Table 19) and radiographic imaging features (Supplementary Table 20).

In contrast to minimal overlap in genome-wide significant results with asthma and smoking, we discovered a significant overall genetic correlation of COPD with asthma (r_genetic = 0.38, P = 6.2×10⁻⁵) using LD score regression in our European-ancestry subjects^46,47. We also assessed genetic correlation with population-based lung function, pulmonary fibrosis, smoking behavior, and two common COPD comorbidities, coronary artery disease and osteoporosis. We identified significant correlation of COPD with lung function and two aspects of smoking behavior, but not with common comorbidities or with pulmonary fibrosis (Figure 4). The lack of significant correlation of COPD with pulmonary fibrosis may indicate our overlapping loci for COPD and pulmonary fibrosis are not representative of a broader disease correlation; alternatively, it could reflect limited sample size or a mix of positive and negative genetic correlations across the genome for the diseases. In potential support of this latter hypothesis, and in contrast to the loci we describe in this study, are recent descriptions of rare variants in telomerase genes predisposing to both emphysema, a key feature of COPD, and pulmonary fibrosis^40,48. Our analysis of partitioned heritability identified COPD genetic association enrichment in fetal lung tissue (coefficient P = 3.5×10⁻⁷); other bioinformatics analyses also support functional annotation of COPD associations to lung tissue or lung cell types (Supplementary Note).

Figure 4. Genetic correlation (using LD score regression) between COPD and other traits.

Shading and numbers represents strength of correlation. An asterisk indicates nominal (P < 0.05) significance, and a double asterisk indicates significant after Bonferroni correction for number of pairwise comparisons. fev1fvc and fev1 = lung function (FEV1/FVC ratio and FEV₁ from CHARGE/SpiroMeta⁴, asthma taken from the asthma GWAS by the GABRIEL Consortium⁴³, ild = pulmonary fibrosis from Fingerlin et al.^8,9, bilSmk = subset of smokers in the UK BiLEVE study⁷, smkCpd = cigarettes per day smoking from the Tobacco and Genetics (TAG) Consortium⁴⁵, smkFormer = current versus former smokers from TAG, smkOnset = age of smoking initiation from TAG, smkEver = ever versus never smoking from TAG. cad = coronary artery disease from the CARDIoGRAM study⁵⁰, height⁵¹ and bmi (body mass index)⁵² from the GIANT consortium, bmdLumbar and bmdFemoral = lumbar and femoral bone mineral density, respectively, from the Genetic Factors for Osteoporosis (GeFOS) Consortium⁵³.

Our study is, to our knowledge, the largest genome-wide association study of COPD cases to date and includes over 60,000 subjects (including 15,256 COPD cases) in the Stage 1 analysis. We chose to combine subjects of different ethnicities, hypothesizing that shared COPD risk factors across ethnicities would outweigh power loss due to heterogeneity. While methods have been developed that can more rigorously assess the degree of overlap and provide additional power in this setting⁴⁹, none of our non-white cohorts were sufficiently sized or powered for these analyses. COPD is also a highly heterogeneous disease; whether a more precise phenotypic definition would result in greater power is not clear. We used a staged study design and examined overall meta-analysis P-values to determine genome-wide significance. Thus, 9 loci (TET2, CFDP1, TGFB2, AGER, ARMC2, PID1, MTCL1, SFTPD, and CYP2A6) from our Stage 1 analysis, which only reached genome-wide significance in either the Stage 2 UK BiLEVE analysis or the overall meta-analysis, should be further replicated. However, six of these 9 association signals are significant if we consider a Bonferroni correction (P < 6.3×10⁻⁴) for the 79 variants tested in Stage 2. Further, 8 of these 9 variants are more strongly associated in the overall meta-analysis compared to Stage 1; the exception is RARB, which has a previously reported association with both lung function⁴ and airflow obstruction²¹ (Table 1).

The majority of our significant loci overlap with lung function loci, strengthening the foundation for investigating the relationship of lung function variability in the general population to risk of developing COPD. These loci are unlikely to reflect susceptibility for asthma or for cigarette smoking; however, our association results as a whole show evidence of shared heritability with asthma (supporting investigation into shared genetic etiologies for these diseases) and cigarette smoking behavior (despite adjustment for smoking in our statistical model). We identified functional annotation enrichment for fetal lung cells, supporting a role for early life events contributing to future risk of COPD. Finally, we identify loci that overlap with pulmonary fibrosis, but with opposite risk alleles. Our study highlights the important contribution of genetic association studies to understanding COPD, not only by identifying novel loci, but also illustrating relationships with other pulmonary traits and diseases.

Data Availability Statement

The genome-wide association summary statistics generated in the Stage 1 analysis of the current study are available in the dbGaP repository, https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000179.v5.p2

The Stage 2 analysis summary statistics are available in Supplementary Table 3.

Online Methods

Study Cohorts

We invited investigators from 22 studies with genome-wide association data and COPD case-control or general population samples with spirometry to participate in a genome-wide association meta-analysis. Additionally, we included four cohorts with Illumina HumanExome v1.2 and custom genotyping based primarily on prior top results from a previously published COPD GWAS¹³, using results with P < 1×10⁻⁴ using plink ‘–clump’ on the COPDGene non-Hispanic whites to perform linkage disequilibrium pruning (r² < 0.8), preferentially retaining both an imputed and genotyped top SNP at each locus. An additional group of variants was a candidate panel, based on results from a previous candidate gene analysis⁵⁴, as well as variants identified in association with lung function (supplementing the existing content on the array, which included variants from previous genome-wide association studies), including the lead SNP and a 200kb region around that SNP pruned for variants with P < 0.01 and r² < 0.8, and additional top-ranked SNPs for COPDGene-specific analyses for lung function, bronchodilator responsiveness, exacerbations, and SNPs from candidate genes.

The baseline characteristics of these 26 cohorts can be seen in Supplementary Table 1. Each cohort obtained approval from appropriate ethical/regulatory bodies; informed consent was obtained for all individuals. (Further cohort-specific methods can be found in the Data Supplement.) As most of these cohorts did not have post-bronchodilator spirometry, we used a modified definition of GOLD criteria based on pre-bronchodilator spirometry: forced expiratory volume in 1 second (FEV₁) < 80% and FEV₁ to forced vital capacity (FVC) ratio of < 0.7 for cases, and FEV₁ > 80% and FEV₁/FVC > 0.7 for controls. Logistic regression was performed in each cohort, adjusting for age, sex, pack-years of smoking, ever-smoking status, current-smoking status, and ancestry-based principal components, as appropriate for each study. Summary statistics were assessed using EasyQC⁵⁵ version 10.1. More detailed cohort information, including cohort-specific methods, can be found in the Supplementary Note.

Genome-wide association quality control

Summary statistics, including effect allele and other allele oriented to the + strand, effect allele frequency, chromosome and position (hg19), and imputation quality were uploaded to a secure site at the Brigham and Women’s Hospital/Channing Division of Network Medicine. Quality control assessments included assessing allele frequencies versus 1000 Genomes reference, standard error versus sample size, and quantile-quantile plots. Variants with an imputation quality metric of < 0.3 (provided a higher threshold for imputation quality was not already implemented), a minor allele count (MAC) of < 20 using the effective sample size or the number of cases and adjusted for imputation quality where applicable, were set to missing. Variants were included for meta-analysis if they were present in at least 13 studies (those with European ancestry and at least 7 million markers passing all quality control filters).

Staged GWAS meta-analysis

In Stage 1 of the analysis, we used Metal^56,57 version 2011-03-25 to perform a fixed-effects meta-analysis of genome-wide data from 22 studies and four additional COPD cohorts genotyped on an Illumina HumanExome v1.2 platform with custom content; this content included a set of COPD candidate genes and regions identified from prior COPD GWAS efforts¹³. We adjusted for inflation using genomic control correction in each study. We included study populations with subjects of non-European ancestry in the overall analysis, and additionally examined results limited to study populations of European ancestry. To identify variants to test for association in Stage 2 in the UK BiLEVE study, we selected top results (P < 5×10⁻⁶) from the Stage 1 meta-analysis. We selected one lead variant from the chromosome 15q25, FAM13, and HHIP regions, as all of these have been described in multiple COPD GWASs^13,16,17,21. For the remainder of the regions, we performed linkage disequilibrium pruning using the PLINK1.9 –clump procedure with an r² of 0.5, additionally examining these SNPs for the number of cohorts with passing quality control at each variant and including SNPs in strong LD (i.e., part of the same clump) with a lower degree of missingness. To identify independent results, we used GCTA-COJO^58,59 on the Stage 1 meta-analysis for variants with P < 5×10⁻⁶ using the default distance of 10Mb. We used the COPDGene non-Hispanic whites (as the largest representative population) as the reference population for these analyses. An overall meta-analysis across the Stage 1 and Stage 2 (UK BiLEVE) cohorts was performed and variants with P < 5×10⁻⁸ were considered genome-wide significant (Figure 1).

Lung eQTL analysis

Lung expression quantitative trait loci (eQTL) were calculated from 1,111 human subjects who underwent lung surgery at three academic sites, Laval University, University of British Columbia (UBC), and University of Groningen, henceforth referred to as Laval, UBC, and Groningen, respectively. This lung eQTL dataset has been described previously^22,60. Briefly, 66.7% to 91.2% of the individuals in this study were current or former smokers and 24.2% to 35.3% had moderate to severe COPD (GOLD spirometry grade 2 to 4). Whole-genome gene expression profiling in the lung was performed on a custom Affymetrix array (GPL10379). Microarray pre-processing and quality controls were described previously^22,61,62. Probe sequences were mapped to the human genome (hg19) using Bowtie⁶³ and probes not mapping to a coding region or having a common SNP (MAF ≥ 5%) in their sequence were removed. Expression data were adjusted for age, sex, and smoking status using residuals obtained with the robust fitting of linear models function (rlm) in the R statistical package MASS. Residual values deviating from the median by more than three standard deviations were filtered as outliers. Genotyping was carried on the Illumina Human 1M-Duo BeadChip array.

Twenty-one out of the 22 SNPs (in main manuscript Table 1) were genotyped or imputed in the three cohorts, i.e. Laval, UBC, and Groningen. One of the SNPs, rs7186831, was not well-imputed; a proxy, rs11865296 in modest linkage disequilibrium (r² = 0.54, 1000 genomes phase 3, EUR) was used instead. These variants were tested for association with adjusted expression traits (43,465 probe sets) in the lung. SNPs within 1 Mb up and downstream of the transcription probe set were considered as local-eQTL. Distant-acting eQTLs were further than 1 Mb away or on a different chromosome. Association tests were carried with PLINK1.9^64,65 in each cohort and then meta-analyzed using Fisher’s method. All local eQTL with nominal P value < 0.05 in the meta-analysis were considered. To provide an additional overall estimate of eQTL significance, we considered a Bonferroni correction threshold ([0.05/(22 SNPs × 43,465 probe sets) = P value < 5.2 × 10⁻⁸]). Statistical analyses were performed in R3.2.3⁶⁶.

Co-localization Analysis

Co-localization of statistical signals between COPD genetic association and eQTL were examined using the coloc R package²³. We used phenotypic summary statistics from individuals of European ancestry with genome-wide association data and all eQTL results and examined 500kb flanks around the top 22 genome-wide significant associations found in the overall meta-analysis (Table 1).

Sensitivity Analysis

To estimate the effect of using pre- instead of post-bronchodilator lung function on our results, we examined the top set of genome-wide significant results in our largest cohorts with both pre- and post-bronchodilator data and densely imputed genotypes (COPDGene NHW and AA, ECLIPSE, NETT-NAS, and Norway/GenKOLS). Since subjects from these cohorts (except for COPDGene) were included based on post-bronchodilator values, including all subjects with COPD based on post-bronchodilator spirometry would lead to larger sample sizes and make comparison of P-values more difficult. Thus, we chose a random sample of post-bronchodilator cases and controls that matched the number of pre-bronchodilator cases and controls. We performed logistic regression using these equal sized set of pre- and post-bronchodilator cohorts, and meta-analyzed the results.

Asthma overlap analysis

We assessed the overlap between our results and known asthma susceptibility loci. We downloaded information on genome-wide significant (P < 5×10⁻⁸) associations with asthma and asthma-related traits including asthma and hay fever, asthma (childhood onset), asthma (corticosteroid response), bronchodilator response in asthma, pulmonary function decline, and severe asthma in the NHGRI-EBI GWAS Catalog⁴². Additionally, we examined top associated variants (which were not genome-wide significant) in the susceptibility to the asthma-COPD overlap syndrome⁶⁷. In all, we assessed the association statistics of 49 unique asthma-associated trait loci across 26 genomic regions in our Stage 1 meta-analysis results. We also examined the asthma association statistics of our top COPD loci from overall meta-analysis using publically available asthma GWAS data from the GABRIEL Consortium⁴³. For COPD loci not present in the GABRIEL Consortium asthma GWAS data, we attempted to examine proxy SNPs in LD (r² > 0.5, 1000 genomes phase 1 CEU) with our top COPD loci.

To examine the genetic correlation⁴⁷ of COPD and asthma over the entire genome, we performed LD score regression⁴⁶ using summary statistics from publically available asthma GWAS data from the GABRIEL Consortium⁴³. For all comparisons using LD score regression, we filtered to HapMap3 variants, limited to European-ancestry subjects with genome-wide data, and filtered on missingness using default parameters in munge_sumstats.py. For the GABRIEL data, we required a variant to be present in at least 35 of the studies.

Smoking behavior overlap analysis

We downloaded information on genome-wide significant (P < 5×10⁻⁸) associations with the traits “nicotine dependence” and “smoking behaviour” in the NHGRI-EBI GWAS Catalog⁴². We assessed these top smoking-associated SNPs in our Stage 1 meta-analysis results. We also assessed overlap of smoking and COPD in the publically available summary statistics from the 2010 Tobacco and Genetics Consortium GWAS⁴⁵. We evaluated our top COPD loci associations from overall meta-analysis with both cigarettes per day and ever-smoking traits. For COPD risk SNPs not directly analyzed in the Tobacco and Genetics Consortium GWAS, we attempted to examine proxy SNPs in LD (r² > 0.5, 1000 genomes phase1 CEU) with our top COPD loci.

To examine the genetic correlation⁴⁷ of COPD and smoking behaviours (cigarettes per day and ever-smoking status) over the entire genome, we performed LD score regression⁴⁶ using summary statistics from our current COPD study as noted above and publically available summary statistics from the 2010 Tobacco and Genetics Consortium GWAS⁴⁵.

Fine mapping analysis

We attempted to determine, at each locus, whether we could identify a potentially causal variant. We performed these analyses using European ancestry subjects alone, and in all subjects with genome-wide data, and excluded variants that were not present in at least 80% of the full sample. We assumed a single causal variant at each locus, examined a +/− 250kb region around the top variant, and calculated approximate Bayes factors using the method of Wakefield⁶⁸ to determine the 95% credible set. While specific trans-ethnic mapping approaches^69,70 can significantly assist in identifying causal loci, we found that the number of non-European samples in our study were likely insufficient to leverage these methods.

Functional enrichment analysis

To identify enriched cell types for our COPD associations, we applied LD score regression to GenoSkyline⁷¹ lung tissue annotations (the default LD score regression annotations collapse lung into the cardiovascular tissue type), as well as cell-type specific annotations from LD score regression⁴⁶. We also performed analysis using only the 22 genome-wide significant loci and tested for enrichment of imputed chromatin marks from ROADMAP using HaploReg 4.1²⁹. Further, we applied a more sophisticated analysis adjusting for local linkage disequilibrium patterns, GoShifter⁷². Finally, we examined overlap with gene expression datasets using SNPsea⁷³.

Additional pulmonary fibrosis and COPD overlap analysis

To further examine overlapping loci for COPD and pulmonary fibrosis, we combined summary statistics from our study and the pulmonary fibrosis GWAS by Fingerlin et al.^8,9 using gwas-pw³⁷.

Causal gene analysis

For the genome-wide significant loci from the overall meta-analysis, we explored potential causative genes at each association locus using the PrixFixe method⁷⁴, assuming co-function of all significant loci. As required by the PrixFixe method, we assured our genome-wide significant loci were present in dbSNP v137⁷⁵ and were represented HapMap⁷⁶ phase III data; for loci not meeting these requirements, proxy SNPs from HapMap phase III were selected based on strongest LD (r²) with index SNP (see Supplementary Table 11 for details of the proxy variant used at each genome-wide significant locus).

COPD comorbidity overlap analysis

We assessed the overlap between our results and two common COPD comorbidities, coronary artery disease and osteoporosis (through bone mineral density traits). We downloaded information on genome-wide significant (P < 5×10⁻⁸) associations with these comorbidities as reported in the NHGRI-EBI GWAS Catalog⁴². We assessed the association statistics of these comorbid trait loci in our Stage 1 meta-analysis results.

Quantitative imaging overlap analysis

To explore the relationship between our top COPD-associated variants and imaging features of emphysema and airway thickness, we queried data from a GWAS of COPD quantitative imaging features⁷⁷. For each genome-wide significant COPD susceptibility locus in our overall meta-analysis, we assessed the corresponding quantitative imaging GWAS effect size, effect direction, and P value for association with the following quantitative imaging traits: %LAA-950 (percentage low attenuation area, using a threshold of −950 Hounsfield units); Perc15 (value of Hounsfield units at the 15th percentile of the density histogram); Pi10 (airway wall area: the value for a hypothetical airway of 10 mm internal perimeter obtained by plotting a regression line of the square root of the airway wall area versus the airway internal perimeter); and WAP (percentage of the wall area compared to the total bronchial area).

Gene set enrichment analysis

As an attempt to minimize false positives in our gene set enrichment analysis, we divided the Stage 1 GWAS cohorts with full genome-wide data into two sets of roughly equal size. We then used i-GSEA4GWAS (http://gsea4gwas.psych.ac.cn/)⁷⁸ for each of the two GWAS data sets to assess enrichment of COPD GWAS loci in BioCarta (http://cgap.nci.nih.gov/Pathways/BioCarta_Pathways) and KEGG⁷⁹ pathways as well as gene ontology (GO) terms^80,81. We first evaluated GO terms and pathways with a false-discovery rate (FDR) less than 5% in both analysis sets and then used a more stringent threshold of FDR < 1% to evaluate overlap of GO term and pathway enrichment in our two analysis sets.