Colocalization of Gene Expression and DNA Methylation with Genetic Risk Variants Supports Functional Roles of MUC5B and DSP in Idiopathic Pulmonary Fibrosis

Abstract

Rationale

Common genetic variants have been associated with idiopathic pulmonary fibrosis (IPF).

Objectives

To determine functional relevance of the 10 IPF-associated common genetic variants we previously identified.

Methods

We performed expression quantitative trait loci (eQTL) and methylation quantitative trait loci (mQTL) mapping, followed by co-localization of eQTL and mQTL with genetic association signals and functional validation by luciferase reporter assays. Illumina multi-ethnic genotyping arrays, mRNA sequencing, and Illumina 850k methylation arrays were performed on lung tissue of participants with IPF (234 RNA and 345 DNA samples) and non-diseased controls (188 RNA and 202 DNA samples).

Measurements and Main Results

Focusing on genetic variants within 10 IPF-associated genetic loci, we identified 27 eQTLs in controls and 24 eQTLs in cases (false-discovery-rate-adjusted P < 0.05). Among these signals, we identified associations of lead variants rs35705950 with expression of MUC5B and rs2076295 with expression of DSP in both cases and controls. mQTL analysis identified CpGs in gene bodies of MUC5B (cg17589883) and DSP (cg08964675) associated with the lead variants in these two loci. We also demonstrated strong co-localization of eQTL/mQTL and genetic signal in MUC5B (rs35705950) and DSP (rs2076295). Functional validation of the mQTL in MUC5B using luciferase reporter assays demonstrates that the CpG resides within a putative internal repressor element.

Conclusions

We have established a relationship of the common IPF genetic risk variants rs35705950 and rs2076295 with respective changes in MUC5B and DSP expression and methylation. These results provide additional evidence that both MUC5B and DSP are involved in the etiology of IPF.

At a Glance Commentary

Scientific Knowledge on the Subject

Common genetic variants have been associated with idiopathic pulmonary fibrosis (IPF), but their functional consequences have not been fully elucidated.

What This Study Adds to the Field

Common IPF genetic risk variants rs35705950 and rs2076295 are associated with changes in MUC5B and DSP expression and methylation, respectively. These results provide additional evidence that both MUC5B and DSP are involved in the etiology of IPF.

While environmental factors play a role in the development of idiopathic pulmonary fibrosis (IPF) (1, 2), genetic risk factors explain a large portion of attributable risk (3) and represent promising approaches to identify disease before irreversible scarring, understand further disease pathogenesis, and identify additional therapeutic targets for this complex and incurable disease (4). Rare variants in telomerase and surfactant gene families have been associated with familial forms of pulmonary fibrosis but are unusual in sporadic cases of IPF (4–8). Common variants in 17 genetic loci have demonstrated genome-wide evidence for association with IPF (9–14).

To develop an integrated understanding of the rare and common variants located in the 10 primary genome-wide association study (GWAS) loci (12), we performed deep targeted resequencing across all 10 loci (3.15 Mb of DNA) in a large population of IPF patients (N = 3,624) and unaffected control subjects (N = 4,442). In that study, we identified 10 common variants that represent the common independent signals in these IPF risk loci and in aggregate account for at least 40% of the risk of IPF (8). Among them, the MUC5B promoter variant, rs35705950, was the strongest genetic risk variant for IPF (8). Questions remain, however, as to which causal biological mechanisms underlie these genetic associations and how identifying these mechanisms can help us understand disease pathogenesis or modify our approach to disease diagnosis and treatment. Our previous work used targeted approaches to assess the effect of the rs35705950 variant and promoter methylation in the regulation of MUC5B gene expression (15). We have more recently shown that the region around the rs35705950 variant functions as a classically defined enhancer subject to epigenetic programming (16). However, more comprehensive analysis of the MUC5B locus and other IPF-associated genetic loci has not been performed.

The Genotype-Tissue Expression (GTEx) project has shown that the majority of genes have regulatory genetic variants (17) and that expression quantitative trait locus (eQTL) mapping combined with co-localization of genetic and eQTL signal is a powerful approach to identify potentially causal genetic risk variants underlying GWAS signals (18). eQTL and co-localization approaches have also been successfully applied to other complex traits using diseased tissue (19, 20). Recently, GTEx also performed a cell type-interaction eQTL (ieQTL) analysis to analyze cell type–specificity of genetic regulation of gene expression across human tissues (21). Because methylation plays a critical role in regulating gene expression, DNA methylation may alter the effect of genetic variants on gene expression through methylation quantitative trait loci (mQTLs) (22, 23). Integration of mQTL and genetic signal at the same locus has also proven successful in prioritizing potential causal variants (22, 23) and mQTLs have previously been co-localized with genetic signal in COPD (24).

To investigate the functional relevance of common genetic IPF risk variants, we have performed genome, transcriptome, and methylome analyses on lung tissue from IPF and control subjects. Here, we report the results of eQTL and mQTL mapping to comprehensively study the effect of genetic variants on local (cis) gene expression and DNA methylation. We also performed co-localization and mediation analysis of eQTL and mQTL with genetic loci to prioritize potential causal risk variants, as well as functional validation of a region in MUC5B containing a novel mQTL by luciferase reporter assays.

Methods

We highlight the key methods in this section. Full methods are available in the online supplement. Count-level transcriptome and DNA methylome datasets are available through the Gene Expression Omnibus under accession GSE175459.

Study Population

Human tissue was collected after appropriate ethical review for the protection of human subjects through the National Heart Lung and Blood Institute (NHLBI)-sponsored Lung Tissue Research Consortium, Interstitial Lung Disease programs at the University of Colorado, National Jewish Health, University of California San Francisco and Vanderbilt University, as well as Committee for Oversight of Research and Clinical Training Involving Decedents for the Lung Donor Program at the University of Pittsburgh.

Genetic Data Processing and Imputation

We applied standard quality checks to the multi-ethnic genotyping array data (8, 12). Ancestral principal components (PCs) were derived by merging overlapping SNPs between our data and 1,000 Genomes samples. Imputation was performed against the Haplotype Reference Consortium v1.1 panel on the Michigan Imputation Server. This resulted in the final set of 7,975,707 SNPs available for further analysis.

Transcriptome Data Processing

mRNA libraries were prepared with TruSeq stranded mRNA library preparation kits (illumina) and sequenced at the average depth of 80M reads on the NovaSeq 6,000 (illumina). RNA paired-end reads were aligned to Ensembl GrCh38 using Kallisto (25). After quality control and filtering, transcriptome data were available on 21,449 genes/58,288 transcripts in 422 samples.

DNA Methylome Data Processing

DNA was bisulfite treated, labeled and hybridized to Illumina Infinium Human MethylationEPIC BeadChip using standard protocols. Illumina signal intensity files were processed using SeSAMe; within-sample normalization with out-of-band probes and dye bias correction were performed (26). After QC and filtering, methylome data were available on 712,229 probes in 547 samples.

Expression and Methylation Quantitative Trait Locus Analyses

To remove the effects of known and unknown batch effects, we used Probabilistic Estimation of Expression Residuals (PEER) (27). PEER factors are surrogate variables that capture technical and demographic batch effects present in the data. PEER models were run adjusting for age, sex, four genetic principal components, and 30 PEER factors. Four PCs from the genetic data were included to appropriately adjust for observed population stratification (see Figure E1 in the online supplement). Comparison of the top 36 Principal component regression analysis and PEER factors (Figure E2) demonstrates that PEER successfully captured confounding attributed to observed technical and demographic batch effects as well as unobserved population stratification. Because PEER adjusted for some of the effects of diagnosis, we ran eQTL and mQTL models separately in controls and IPF cases. eQTL and mQTL permutation models, in which the most significant SNP is identified for each transcript/CpG, were run using a distance cutoff between each SNP-gene/CpG pair of 1 Mb (2 Mb for the FAM13A locus because of the longer distance between the lead SNP and FAM13A gene [8]) in FastQTL (28). β-distributed P values from FastQTL output were adjusted to a 5% false discovery rate (FDR) using the Benjamini-Hochberg procedure (29). We performed confirmatory mQTL analysis using nominal model testing of all CpGs within 1 Mb (2 Mb for the FAM13A locus) of the lead SNPs in each IPF genetic locus. We examined quantile–quantile plots for all eQTL/mQTL models to determine the extent of genomic inflation; nominal mQTL models did not exhibit any inflation while some inflation was present in all permutation-based models (Figure E3). This is expected due to the nature of permutation-based testing in which case the most significant transcript/CpG-SNP pair is selected for each transcript/CpG tested. In general, residual inflation has been observed in studies of DNA methylation on Illumina arrays (30, 31).

Luciferase Assays in Cultured Cells

MUC5B reporter constructs were amplified (Table E1) from genomic DNA by PCR, cloned into pCR2.1-TOPO (Invitrogen), and subsequently ligated into the pGL3-Promoter vector (Invitrogen). A549 cells were transfected with reporter constructs using Lipofectamine 2,000 (Life Technologies) as instructed by the manufacturer. Approximately 24 hours later, luciferase activity was assayed using the Dual-Luciferase Reporter Assay System from Promega as described previously (32). Firefly luciferase activity was normalized to that of a Renilla luciferase internal control (pRL-SV40; Promega).

Results

Expression Quantitative Trait Loci (eQTLs)

As expected, the IPF subjects were older, more often male, and have a more extensive cigarette smoking history than healthy controls (Table 1). We did not observe significant differences in self-reported race and ethnicity between the groups. However, we did observe differences in the method used to obtain lung tissue between IPF subjects and controls (Table 1).

Table 1.

Cohort Characteristics for Controls and IPF Cases Included in the Gene Expression and DNA Methylation Datasets

Gene Expression Dataset	Control (N = 188)	IPF (N = 234)	P Value
Age	55.3  ± 16.8	61.4 ± 7.5	9.92 × 10−6
Sex (M/F)	110/78 (59/41%)	184/50 (79/21%)	1.11 × 10−5
Self-reported race, White (yes/no)	155/33 (82/18%)	200/34 (85/15%)	0.42
Self-reported ethnicity, non-Hispanic (yes/no/unknown)	162/6/20 (86/3/11%)	197/20/16 (84/9/7%)	0.20
Cigarette smoke (ever/never/unknown)	86/89/13 (46/47/7%)	140/75/19 (60/32/8%)	5.7 × 10−3
Tissue sampling method (biopsy/explant/autopsy/unknown)	4/184/0/0 (3/73/0/1%)	18/145/3/68 (8/62/1/29%)	2.2 × 10−16

DNA Methylation Dataset	Control (N = 202)	IPF (N = 345)	P Value
Age	54.8 ± 16.3	61.6 ± 7.8	7.22 × 10−8
Sex (M/F)	123/79 (61/39%)	263/82 (76/24%)	2.09 × 10−4
Self-reported race, White (yes/no)	168/34 (83/17%)	295/50 (85/15%)	0.46
Self-reported ethnicity, non-Hispanic (yes/no/unknown)	174/8/20 (86/4/10%)	288/25/32 (84/7/9%)	0.30
Cigarette smoke (ever/never/unknown)	90/98/14 (45/48/7%)	209/106/30 (60/31/9%)	1.90 × 10−4
Tissue sampling method (biopsy/explant/autopsy/unknown)	4/193/0/5 (2/96/0/2%)	37/185/2/121 (11/54/1/35%)	2.2 × 10−16

Definition of abbreviation: IPF = idiopathic pulmonary fibrosis.

P values were determined by two tailed Student’s t test for age and Fisher Exact test for all other variables

Using residuals from the PEER model, we ran permutation-based transcriptome-wide cis-eQTL analysis in controls and IPF cases separately. This analysis identified 4,745 transcripts with significant association of genetic variants and gene expression (eQTLs) in controls and 6,047 in cases (FDR-adjusted P < 0.05; Table E2). Of the 20,704 tested transcripts, 16,685 also appeared in GTEx lung eQTLs. Concordance of our results with GTEx, estimated via a ${\widehat{\text{π}}}_1$ calculation, suggested broad replication of our eQTLs in the GTEx dataset (${\widehat{\text{π}}}_1=0.58$ = 0.58). Concordance was particularly high for highly significant eQTLs (q < 5 × 10⁻⁴) as demonstrated by the high correlation coefficient of the effect size (slope from the linear model) in our data compared with GTEx lung eQTL data (Pearson r = 0.90, Figure E4).

We next focused on cis-eQTLs in the 10 IPF genetic loci analyzed in our recent resequencing study (8). Of the 161 transcripts within the boundaries of the 10 IPF genetic loci, defined in Moore and colleagues (8), 27 eQTLs were significant in controls and 24 in cases (FDR-adjusted bpval < 0.05; Table 2). Among these eQTLs, only expression of MUC5B and DSP mRNA was associated with the lead genetic variants identified in the two loci (11p15/MUC5B and 6p24/DSP, respectively). In the MUC5B locus, the presence of the alternate allele (T) compared with the major allele (G) at rs35705950 was associated with higher expression of MUC5B mRNA in controls (Figure 1A, left panel). The effect is not apparent in cases with transcript level without adjustment for PEER factors plotted (Figure 1A, left panel) but it is statistically significant although both significance level and the effect size of the eQTL are diminished in cases compared with controls (Table 2). When PEER-normalized transcript levels are plotted, the effect of rs35705950 on MUC5B expression is visually apparent (Figure E5). In the DSP locus, the presence of the alternate allele (G) compared with major allele (T) at rs2076295 was associated with lower expression of DSP mRNA in controls with diminished effect in cases (Figure 1A, right panel and Figure E5).

Table 2.

Significant eQTLs within the Boundaries of IPF Genetic Loci in Controls and Cases

Controls
IPF Locus (Chr: Region)	Gene Name	SNP ID	Distance, bp	Slope	bpval	adj bpval
3:169723712-169910112	LRRC31	rs9290376	27278	−0.41834	0.006212	0.019114
4:88684849-89177649	NAP1L5	rs2860500	30003	0.304358	0.017387	0.044866
6:7493767-7687767	DSP	rs2076295	21423	−0.66886	2.18E-26	3.49E-25
6:7493767-7687767	AL031058.1	rs2076295	22547	−0.43322	5.25E-08	2.71E-07
7:99889391-100167139	MBLAC1	rs139788064	381983	0.910552	1.89E-05	8.17E-05
7:99889391-100167139	TRIM4	rs2527925	5056	0.651133	5.06E-28	1.09E-26
7:99889391-100167139	AZGP1	rs2527889	−15304	−0.56105	7.53E-08	3.76E-07
7:99889391-100167139	AP4M1	rs9649220	−59212	−0.28104	0.000245	0.000853
7:99889391-100167139	LAMTOR4	rs77712869	−66934	−0.64989	9.85E-05	0.000358
7:99889391-100167139	TAF6	rs1050542	133	−0.17539	6.03E-05	0.000235
7:99889391-100167139	CNPY4	rs144305120	28348	−0.38747	0.010134	0.028445
7:99889391-100167139	ZSCAN21	rs3736591	103619	−0.26315	0.010567	0.029149
10:103837373-104069233	SH3PXD2A	rs11191773	106615	−0.21871	0.005672	0.017795
10:103837373-104069233	SLK	rs2864020	572109	−0.15223	0.007512	0.021464
11:1008000-1737755	MUC5B	rs35705950	−3076	0.560129	2.97E-11	1.76E-10
11:1008000-1737755	KRTAP5-AS1	rs1809668	27323	−0.38391	1.87E-05	8.17E-05
11:1008000-1737755	AC068580.4	rs8839	18084	−0.51894	9.62E-05	0.000358
11:1008000-1737755	IFITM10	rs72843946	136930	−0.43939	0.007374	0.021452
11:1008000-1737755	LINC02688	rs1127800	−179414	0.652602	0.0167	0.044479
13:112648686-113136861	AL356740.1	rs4907571	−654	0.312598	0.002177	0.006966
15:40293756-40521315	INAFM2	rs934935	19311	−0.27894	0.001665	0.00555
15:40293756-40521315	KNSTRN	rs76277863	−25060	1.02178	1.59E-25	2.32E-24
15:40293756-40521315	BAHD1	rs6492945	−41976	−0.19031	0.014081	0.038186
15:40293756-40521315	IVD	rs8033938	5119	−0.62869	8.64E-12	5.32E-11
15:40293756-40521315	DISP2	rs11636147	70693	0.94725	1.07E-10	6.14E-10
19:4635088-4766379	MYDGF	rs72620526	51891	0.211868	0.016958	0.044479
19:4635088-4766379	DPP9	rs62115083	548680	0.215823	0.017666	0.044866

Cases
IPF Locus (Chr: Region)	Gene Name	SNP ID	Distance	Slope	bpval	adj bpval
3:169723712-169910112	LRRC31	rs2276718	30773	−0.31134	0.007675	0.020813
4:88684849-89177649	NAP1L5	rs8605	320	0.233498	0.003354	0.009937
6:7493767-7687767	DSP	rs2076295	21423	−0.15647	0.005279	0.015081
7:99889391-100167139	TRIM4	rs2527926	2942	0.627493	4.29E-48	2.29E-46
7:99889391-100167139	AZGP1	rs2527898	−30314	−0.32512	0.018054	0.045136
7:99889391-100167139	ZSCAN21	rs4729571	869	−0.32981	3.43E-07	1.57E-06
7:99889391-100167139	CNPY4	rs114198920	−54504	−0.61142	8.66E-06	3.08E-05
7:99889391-100167139	AP4M1	rs13236456	−23615	−0.23271	1.58E-05	5.38E-05
7:99889391-100167139	TAF6	rs2293481	−5068	−0.22596	4.13E-06	1.57E-05
7:99889391-100167139	MBLAC1	rs146579476	6792	0.970627	2.41E-08	1.13E-07
11:1008000-1737755	MUC5B	rs35705950	−3076	0.307958	8.32E-06	3.03E-05
11:1008000-1737755	TOLLIP-AS1	rs80158222	−85073	−0.43748	1.62E-06	6.33E-06
11:1008000-1737755	BRSK2	rs36064646	−935	−0.26786	0.0198	0.048739
11:1008000-1737755	KRTAP5-AS1	rs34995599	9984	−0.42908	2.21E-11	1.26E-10
11:1008000-1737755	IFITM10	rs111693235	16583	−0.27613	0.010576	0.02774
11:1008000-1737755	AC068580.4	rs111693235	14172	−0.26365	0.014982	0.038049
13:112648686-113136861	AL139384.1	rs2274774	2122	−0.26887	5.40E-06	2.01E-05
13:112648686-113136861	AL356740.1	rs4907571	−654	0.428474	1.81E-11	1.07E-10
15:40293756-40521315	CHST14	rs28521889	−249578	0.492508	0.005897	0.016554
15:40293756-40521315	INAFM2	rs7171143	−195	−0.3379	1.09E-05	3.79E-05
15:40293756-40521315	DISP2	rs56221586	−35	0.975899	3.55E-18	2.71E-17
15:40293756-40521315	KNSTRN	rs17671194	−7384	1.13297	1.25E-39	3.99E-38
15:40293756-40521315	IVD	rs8033938	5119	−0.52893	2.52E-12	1.55E-11
19:4635088-4766379	DPP9	rs12462642	39779	−0.22049	0.003044	0.009188

Definition of abbreviations: adj bpval = Benjamini-Hochberg adjusted bpval; bp = base pairs; bpval = P value of association adjusted for the number of variants tested in cis given by the fitted β distribution; Chr = chromosome; eQTLs = expression quantitative trait loci; IPF = idiopathic pulmonary fibrosis.

Highlighted are eQTLs associated with the lead variants from the IPF resequencing study by Moore et al. (8)

Figure 1.

eQTLs at MUC5B and DSP loci co-localize with genetic signal. (A) Box plots for MUC5B and DSP eQTLs. Normalized RNA TMP refers to transcripts per million after trimmed mean of M values (TMM) normalization across samples and inverse normal transformation on a per-gene basis, as has been done in GTEx. eQTL analysis was performed in 188 controls and 234 cases with the following genotype breakdown. MUC5B rs35705950: 141 GG, 42 GT, and 5 TT in controls and 105 GG, 114 GT, and 15 TT in IPF. DSP rs2076295: 54 TT, 90 GT, and 44 GG in controls and 54 TT, 101 GT, and 79 GG in IPF. (B) Mirror plots for co-localization of eQTL (top) with genetic signal (bottom) at MUC5B and DSP loci. CLPP = co-localization posterior probability; eQTL = expression quantitative trait loci; IPF = idiopathic pulmonary fibrosis.

We also performed a subgroup analysis of cis-eQTLs in the 10 IPF genetic loci in 140 controls and 140 cases frequency-matched on age, sex, and smoking (Table E3). Given the reduction in sample size, it is not surprising that eQTLs in MUC5B and DSP remained highly significant in controls but did not reach significance in cases. However, the effect size measured by the slope in the linear model remained similar. For MUC5B, slopes are 0.560 in the full cohort and 0.553 in the subgroup analysis in controls, and 0.308 versus 0.247 in full cohort versus subgroup analysis in cases. For DSP, slopes are -0.669 in the full cohort and -0.695 in the subgroup analysis in controls, and −0.156 versus −0.166 in full versus subgroup analysis in cases. We have previously observed less pronounced effects of the MUC5B genetic variant on lung gene expression in cases than in controls (15, 33) and the reduced effect size in cases is in line with these observations. We believe that extensive remodeling of IPF lung that is associated with changes in expression of many genes may be responsible for masking some of the effects of the genetic variants compared with what we observed in controls. Overall, however, this subgroup analysis demonstrates that age, sex and smoking do not significantly influence our key eQTL findings.

Co-localization of eQTLs and Genetic Loci

To assess potential causality of the genetic variants that act as eQTLs within the IPF loci, we performed a co-localization analysis of genetic (from the resequencing study; exclusively non-Hispanic White [8]) and eQTL (from the current study; >80% non-Hispanic White) data. Using eCAVIAR co-localization posterior probability (CLPP) scores, we demonstrate almost perfect co-localization of genetic and eQTL signal for the MUC5B promoter variant rs35705950 (CLPP = 1.00 in controls and 0.96 in cases; Figure 1B, left panel). We also observed a strong co-localization of genetic and eQTL signal for the DSP risk variant rs2076295 (CLPP = 1.00 in controls and 0.66 in cases; Figure 1B, right panel). Examination of all CLPP scores (Table E4) revealed only a few other co-localization results in the chr11 and chr6 loci. rs12802931 and MUC5B had CLPP of 0.93 in controls and 0.50 in cases; this variant is in linkage disequilibrium with rs35705950 (D’ = 1.00, R² = 0.54 in the EUR population), indicating that rs12802931 does not represent an independent signal. Moreover, conditional testing at rs35705950 indicated that rs12802931 does not represent an independent signal from that of rs35705950 (P = 0.27 after conditional testing). On chr6, rs2076295 had a CLPP of 0.99 in controls with lncRNA AL031058.1, which is antisense to DSP and located in the promoter of DSP. However, we observed no evidence of co-localization in cases (CLPP = 0.040). Similarly, rs55938083 has a CLPP of 0.89 with DSP and 0.44 with lncRNA AL031058.1 in controls but this co-localization was not present in cases (CLPP = 6.3 × 10⁻⁴ for DSP and 9.4 × 10⁻³ for lncRNA AL031058.1). This variant is in linkage disequilibrium with rs2076295 (D’ = 1.00, R² = 0.54 in the EUR population) and therefore not an independent signal. All other CLPP scores in controls and cases were low (CLPP < 0.10) and provided no evidence for co-localization of genetic and eQTL signals in the remaining IPF loci.

Cell Type Interaction eQTLs (ieQTLs)

To address the issue of cell type-specificity in complex tissue, we performed cell type–interaction eQTL (ieQTL) analysis restricted to IPF loci in seven cell populations that are enriched in lung tissue (Figure E6A). This ieQTL analysis revealed 81 significant associations in total, ranging between 1 and 7 significant ieQTLs in each of the 14 models (7 cell types, controls and IPF cases; Table E5). Only one of these associations is related to an IPF locus lead variant, rs2076295 with DSP in smooth muscle cells. We observed decreased expression of DSP with the alternate allele, as expected from our eQTL analysis and previous work (12), when enrichment for smooth muscle cells is low; the effect of the allele on expression diminishes with increasing enrichment of smooth muscle cells (Figure E6B). eCAVIAR analysis demonstrated co-localization of this smooth muscle cell ieQTL in controls with the genetic signal (Figure E6B; CLPP = 0.87). As ieQTLs may capture compensatory cell abundance changes, we asked whether this anticorrelation with smooth muscle cell enrichment may be driven by a compensatory increased abundance in another cell type. Because of the interest in DSP expression in epithelial cells in IPF (34), we also visually examined the effect in epithelial cells, even though no significant interaction was observed; opposite of smooth muscle results, the alternate allele is associated with a decrease in DSP expression in controls when enrichment for epithelial cells is high. These data suggest that the effect of the genetic variant on DSP expression in control tissue is potentially attributable to epithelial cells. However, epithelial cell populations in lung tissue are complex and these preliminary results need further validation. We did not observe any ieQTL effects in IPF lung tissue, likely because of the widespread aberrant expression of DSP in diseased lung, including epithelial cells (Figure E7).

DNA Methylation Quantitative Trait Loci (mQTLs) and Co-localization of mQTLs and Genetic Loci

We next focused on the possibility that IPF genetic variants may influence DNA methylation. We performed both a permutation-based analysis of all SNP-CpG pairs within the boundaries of the risk loci (analogous to the analysis we performed for the eQTLs) and nominal analysis of only the 10 lead risk variants (as previously published mQTL analyses used nominal models [35–40]). Permutation analysis revealed on the order of 10 to a few hundred significant associations of the genetic variants with DNA methylation levels (cis-mQTLs) in each genetic locus (Table 3 and Table E6); of these, seven mQTLs in controls and four mQTLs in cases were associated with lead genetic variants, and only two (rs35705950-cg17589883 in the MUC5B locus and rs2076295-cg08964675 in the DSP locus) were identified in both controls and cases (Table 4). More conventional nominal mQTL models focusing on only lead genetic variants confirmed the association of these two SNP-CpG pairs (Table E7). We also performed subgroup permutation and nominal mQTL analyses in the same set of 140 controls and 140 cases frequency-matched on age, sex, and smoking that were used in the subgroup eQTL analysis. Not surprisingly, this reduction in cohort size substantially reduced the number of significant loci both in permutation and nominal models (Table E8). Importantly, however, the effect sizes remain similar. For rs35705950-cg17589883 mQTL in the MUC5B locus, slopes are 0.201 in the full cohort and 0.205 in the subgroup analysis in controls, and 0.142 versus 0.166 in full versus subgroup analysis in cases. For rs2076295-cg08964675 mQTL in the DSP locus, slopes are 0.223 in the full cohort and 0.196 in the subgroup analysis in controls, and 0.152 versus 0.149 in full versus subgroup analysis in cases. These results demonstrate that age, sex, and smoking do not significantly influence our key mQTL results.

Table 3.

Significant mQTLs within the Boundaries of IPF Genetic Loci. Summary of Significant mQTLs in Permutation-based and Nominal Analyses

IPF Locus (Chr: Region)	Number of CpGs Tested	Significant mQTLs in Controls	Significant mQTLs in Cases
3:169723712-169910112	609	15	20
4:88684849-89177649	816	33	38
5:1213146-1385099	2047	37	80
6:7493767-7687767	618	31	31
7:99889391-100167139	1687	51	95
10:103837373-104069233	875	41	60
11:1008000-1737755	3347	221	348
13:112648686-113136861	2256	178	244
15:40293756-40521315	1272	70	103
19:4635088-4766379	1753	12	15

Definition of abbreviations: Chr = chromosome; IPF = idiopathic pulmonary fibrosis; mQTLs = methylation quantitative trait loci.

Table 4.

Significant mQTLs Associated with Lead Variants in IPF Loci from the IPF Resequencing Study by Moore et al. (8)

SNP	CpG	Gene Name	CpG Gene Relation	Distance, bp	Slope	bpval	Adjusted bpval
Controls
rs2609260	cg25055244	FAM13A	Body	48245	−0.13281	0.00292	0.044863
rs2076295	cg08964675	DSP	Body	423	0.222649	3.02E-05	0.000989
rs2076295	cg12817734	DSP	Body	9397	−0.07546	8.18E-08	5.05E-06
rs35705950	cg03298405	MUC5B	Promoter	−1206	0.133362	0.000937	0.017711
rs35705950	cg16842717	MUC5B	Body	−11050	0.329679	0.000101	0.002513
rs35705950	cg17589883	MUC5B	Body	−18978	0.200628	4.44E-07	1.60E-05
rs35705950	cg19488922	MUC5B	Body	−8223	0.237018	0.000579	0.011594
IPF cases
rs2076295	cg03818715	SNRNP48	Body	−28116	0.091787	0.000559	0.01329
rs2076295	cg08964675	DSP	Body	423	0.151517	0.000201	0.005579
rs35705950	cg02522041	MUC5B	Body	−7787	0.07022	0.000376	0.004949
rs35705950	cg17589883	MUC5B	Body	−18978	0.142382	6.92E-08	1.48E-06

Definition of abbreviations: adj bpval = Benjamini-Hochberg adjusted bpval; bp = base pairs; IPF = idiopathic pulmonary fibrosis; mQTLs = methylation quantitative trait loci.

Highlighted are eQTLs associated with the lead variants from the IPF resequencing study by Moore et al. (8)

The presence of the alternate allele (T) compared with the major allele (G) at rs35705950 was associated with higher methylation at the cg17589883 CpG within the gene body (exon 26) of MUC5B in both controls and cases (Figures 2A and E8, left panel). Similarly, the presence of the alternate allele (G) compared with major allele (T) at rs2076295 was associated with higher methylation at the cg08964675 CpG within the gene body (intron 4) of DSP in both controls and cases (Figures 2A and E8, right panel). Both effects were more pronounced in controls than cases, analogous to our eQTL results. Similar to the eQTL co-localization, we observed almost perfect co-localization of genetic and mQTL loci for rs35705950 and cg17589883 (CLPP of 0.97 in controls and 0.99 in cases; Figure 2B, left panel). We also observed a strong co-localization of genetic and mQTL loci for rs2076295 and cg08964675 (CLPP of 0.70 in controls and 0.89 in cases; Figure 2B, right panel).

Figure 2.

mQTLs at MUC5B and DSP loci co-localize with genetic signal. (A) Box plots for MUC5B (cg17589883) and DSP (cg08964675) mQTLs. Normalized methylation β refers to β values of DNA methylation level measured on Illumina arrays (scale 0–1) after SeSame data preprocessing and normalization. mQTL analysis was performed in 202 controls and 345 cases with the following genotype breakdown. MUC5B rs35705950: 165 GG, 35 GT, and 2 TT in controls and 160 GG, 164 GT, and 21 TT in IPF. DSP rs2076295: 61 TT, 89 GT, and 52 GG in controls and 88 TT, 149 GT, and 108 GG in IPF. (B) Mirror plots for co-localization of mQTL (top) with genetic signal (bottom) at MUC5B and DSP loci. CLPP = co-localization posterior probability; IPF = idiopathic pulmonary fibrosis; mQTL = methylation quantitative trait loci.

Mediation Analysis

Given the evidence for potential causality at the MUC5B and DSP genetic risk loci from co-localization analyses, we performed a mediation analysis by running a series of regression models, while adjusting for sex and four genetic PCs (results of all models summarized in Table E9). The association of rs35705950 with disease risk was attenuated from an odds ratio (OR) of 3.08 (95% confidence interval [CI], 1.96–4.85; P = 1.14 × 10⁻⁶) to 1.95 (95% CI, 1.18–3.21; P = 8.74 × 10⁻³) after adjusting for MUC5B transcript expression (36.7% reduction in OR; 95% CI, −49.55% to −21.99%). Conversely, the association of rs2076295 with disease risk was enhanced from an OR of 1.14 (95% CI, 0.84–4.27; P = 0.39) to 1.77 (95% CI, 1.22–8.54; P = 2.85 × 10⁻³) when DSP transcript expression is included in the model (54.8% increase in OR; 95% CI, 26.04–107.58%). Because the presence of the alternate allele at rs35705950 is positively associated with both MUC5B expression and disease risk, the attenuation of the association between rs35705950 and IPF is suggestive of partial mediation of the effect of the genetic variant on disease risk. Because the alternate (G) allele at rs2076295 is negatively associated with DSP expression but positively associated with disease risk, the observed enhancement of the effect of the G allele on disease risk after adjustment for expression is consistent with expression being a negative confounder in the relationship between the allele and disease risk. We have noted this complex relationship in the DSP locus previously (12). We did not observe any evidence of mediation through DNA methylation (Table E10).

Transcriptional Activity of the MUC5B Locus

Examination of genomic functional element annotation provided by The Encyclopedia of DNA Elements (ENCODE) Consortium indicates that both CpGs reside in regions enriched for transcription factor binding, sensitivity to nuclease digestion, and histone marks associated with open/active chromatin (Figure E9), and therefore may be important in gene regulation. To determine whether the region in the MUC5B locus harboring the cg17589883 CpG has a functional role, we generated a luciferase reporter containing this CpG and several hundred base pairs of flanking DNA (referred to as “Reg 4”; see Figure 3). We also generated reporters from several additional putative regulatory regions of the MUC5B locus based on ENCODE data. We tested these reporters along with a previously described construct spanning the MUC5B rs35705950 site that exhibits enhancer activity (“Reg 1”) (16). In this assay, the cg17589883 CpG region repressed reporter activity in the “Reg 4” construct, whereas a region near the 3′ UTR in the “Reg 6” construct functioned as an enhancer (Figure 3). The other tested regions had minimal effect. These data provide further support for a functional role of the CpG-containing region in regulating MUC5B expression. Future work will need to determine whether this repressor region within the MUC5B gene interacts with the enhancer containing rs35705950.

Figure 3.

Reporter assays ascribe selective regulatory function to cg17589883 region in MUC5B. (A) UCSC Genome Browser screenshot of the MUC5B locus with H3K27Ac Mark, DNase I Hypersensitivity Peak, and Transcription Factor (TF) ChIP-seq Cluster tracks from The Encyclopedia of DNA Elements Consortium enabled and used to identify regions with putative regulatory function for cloning into luciferase reporters, as indicated. (B) Mean ± SD normalized luciferase activity of indicated reporter constructs and empty vector (EV) control in A549 cells under basal culture conditions (n = 4 per group and *P < 0.0001 versus EV via one-way ANOVA with Bonferroni correction). ANOVA = Analysis of Variance; Reg = region.

Discussion

Our findings lead us to conclude that the IPF risk variants rs35705950 and rs2076295 likely perturb the expression and methylation of MUC5B and DSP, respectively, and are involved in the etiology of IPF. More specifically, our findings demonstrate strong co-localization of rs35705950 and rs2076295 with both gene expression and DNA methylation marks. Moreover, mQTL results led to identification of a putative internal repressor element within MUC5B. Collectively, these results provide additional evidence that both MUC5B and DSP are involved in the etiology of IPF, and that the expression of these genes is regulated by both genetic and epigenetic factors. However, using transcriptional and epigenetic approaches, we were unable to identify potential regulatory roles for the other eight common IPF risk variants (8) and further work will be necessary to disentangle the causal variants and perturbed biological pathways at these loci.

Several lines of investigation, in addition to the findings presented in this manuscript, suggest that MUC5B is involved in the etiology of pulmonary fibrosis. First, rs35705950 (15, 33) is the dominant risk factor for IPF and is present in >50% of affected patients (33); this finding has been validated in at least 10 independent studies (8, 9, 12, 14, 33, 41–47). Second, MUC5B is normally not expressed in the terminal bronchioles (48–50). However, MUC5B is expressed in the bronchiolar epithelia, epithelial cells lining honeycomb cysts, and co-expressed with surfactant protein C (SFTPC) in alveolar type II cells in IPF (33, 51–54), indicating that cell types involved in lung fibrosis in the distal airspace express MUC5B. In addition, the variant allele is specifically associated with increased expression of MUC5B in the terminal bronchiole (52). Third, overexpression of Muc5b in bronchoalveolar epithelia in mice is directly related to the extent and persistence of bleomycin-induced lung fibrosis, honeycomb metaplasia, and mortality (53, 55). Fourth, we recently demonstrated that the MUC5B variant rs35705950 resides within an enhancer that is subject to epigenetic remodeling and contributes to pathologic misexpression in IPF (16). Findings presented in the current study demonstrate further the importance of the MUC5B promoter variant on both risk and MUC5B expression and identify the regulatory importance of methylation marks in repressing the effects of the MUC5B promoter variant.

Accumulating evidence suggests that DSP, part of the desmosome, is also involved in the etiology of IPF. DSP is critical to cell–cell adhesion, wound repair, and epithelial barrier function. While rs2076295 has been repeatedly found to be associated with IPF (8, 34), other variants in DSP have also been associated with cardiac fibrosis (56), right ventricular dysplasia (57), and keratodermas (58), suggesting that loss of cell–cell adhesion may result in a number of conditions involving injury, tissue remodeling, and fibrosis. Deletion of the DNA region spanning rs2076295 using CRISPR/Cas9 editing leads to reduced expression of DSP and an edited G allele at rs2076295 results in lower expression of DSP compared with the wild-type T allele in a human bronchial epithelial cell line (59). This is consistent with the results of our eQTL analysis that show a decrease in DSP gene expression with the presence of the alternate allele. Moreover, loss of DSP enhanced ECM-related gene expression and promoted cell migration, potentially contributing functionally to the pathogenesis of IPF (59). A recent study also showed that stiff matrix induces DSP gene expression in lung epithelial cells and that this induction is regulated by DNA methylation of a conserved region in the proximal DSP promoter (60). In aggregate, these findings suggest that reduced expression of DSP could adversely affect wound healing and promote fibroproliferation.

Our findings underscore the complex etiology of IPF. The presence of both genetic and epigenetic changes associated with disease-defining transcriptional changes in the IPF lung strongly suggest that lung fibrosis is driven by gene-by-environment interactions. Although the MUC5B promoter variant is the dominant genetic risk variant for the development of IPF (33), chronic hypersensitivity pneumonitis (61), rheumatoid arthritis–associated interstitial lung disease (62), and asbestosis (63), only a small portion of individuals with these common genetic variants go on to develop lung fibrosis, raising the possibility that environmental factors such as microscopic lung injury caused by inhaled particles or toxins are inadequately cleared and/or cause excessive lung injury in genetically susceptible hosts. In fact, several genetic studies (33, 64–67) indicate that a specific gene variant or locus may cause different types of lung fibrosis within the same family. This supports our hypothesis that sharing the same genetic variant does not necessarily result in the same disease pattern and that other influences, including specific environmental exposures or possibly other genetic variants are pivotal to the final phenotype that emerges.

It is particularly intriguing that of the 10 IPF common risk loci, only MUC5B and DSP co-localized with gene expression and methylation marks. Importantly, variants in telomerase genes may exert their effects through circulating cells or at different stages of lung fibrosis; TERT and TERC have undetectable expression in lung tissue and therefore were not included in the analysis and lead genetic variants at the OBFC1/STN1 locus had no eQTLs. Genetic variants in telomerase genes are much more likely to function by telomere shortening (68) or cell senescence in alveolar cells (69). It is also important to note that this study focused on cis-eQTLs in the 10 IPF common risk loci identified by our group but that we report genome-wide cis-eQTL results for all genotype-expression pairs and that these data can be mined by others for additional IPF common variant loci.

The main limitation of the current study is the use of whole lung tissue. Cell heterogeneity in whole lung tissue is likely the reason we observe small effect sizes in our mQTL results. We partially addressed the issue of cell heterogeneity by performing cell type-interaction eQTL analysis, as has been done by GTEx (21). However, a major limitation of this analysis is that deconvolution of whole lung tissue gene expression data does not work well on specialized cell types in the lung such as MUC5B-producing secretory cells. Future single cell eQTL/mQTL studies (70) will be necessary to fully address cell specificity of the relationship of genetic variant to gene expression and DNA methylation. Another limitation that is inherent to using bisulfite conversion for DNA methylation measurements is inability to distinguish 5-methylcytosine from 5-hydroxymethylcytosine. However, given that the relative amount of 5-hydroxymethylcytosine is small compared with 5-methylcytosine, this is only a minor limitation and likely does not influence the key findings. Lastly, sample size and power were limited in our subgroup analysis of age-, sex-, and smoking-matched cases and controls. Despite these limitations, our study provides substantial evidence for roles of rs35705950 (MUC5B) and rs2076295 (DSP) genetic variants in the development of IPF. Future studies should also use additional assays to assess chromatin accessibility (ATAC-seq, for example) and tools to assess the functionality of the identified eQTL/mQTLs (CRISPR-Cas9, for example) in primary cells.