GWAS for Mucin pQTL

We conducted a GWAS of total and specific (MUC5AC and MUC5B) mucin concentrations in sputum to identify novel regulators of these biomarkers in SPIROMICS. Descriptive statistics of study participants are provided in Table 1. The mucin phenotype data represent a subset of subjects described in two previous studies [8,9], and comprise a subset of participants in SPIROMICS (S1 Fig). In this sample, there was a clear effect of smoking history on total mucin concentration (S2 Fig), but among COPD cases, there was not a linear or monotonically increasing relationship between total mucin concentration and disease severity (as reflected by Global Initiative for Chronic Obstructive Lung Disease (GOLD) stage). Similar patterns were observed for MUC5AC and MUC5B (S2 Fig).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive Statistics of SPIROMICS Participants in Mucin GWAS.^*

https://doi.org/10.1371/journal.pgen.1010445.t001

We did not detect any genome-wide significant loci (p<5.0x10^-8) associated with total mucin concentration based on data from SPIROMICS participants of European ancestry (EA, N = 576) or African ancestry (AA, N = 132) participants (S3 and S4 Figs), nor in combined analysis of EA and AA subjects (S5 Fig). Testing for joint effects of SNP and SNP × smoking (pack-years) interactions did not reveal any loci associated with total mucin concentration either. In contrast, despite relatively small sample size (N = 215 EA subjects), we identified three genome-wide significant pQTL for MUC5AC or MUC5B (Figs 1, S6 and S7 and S1 Table). In addition to the pQTL for MUC5AC on chromosome (Chr) 7 (rs75401036), we identified one highly suggestive locus on Chr 2 (rs16866419, p = 7.2x10^-8), thus both MUC5AC pQTL are located on chromosomes other than Chr 11 where MUC5AC and MUC5B are located (i.e., act in trans). We note that restricting our analysis to variants located in/near MUC5AC, i.e., with a reduced multiple testing correction, did not reveal any local pQTL for MUC5AC, and this includes testing variants previously associated with MUC5AC gene expression [10–12] including rs12788104, rs11602802, rs11603634, rs1292198170, and rs1132436. Overall, SNP-based heritability estimates for MUC5AC (h²_SNP = 0.712; S.E. = 1.322) and MUC5B (h²_SNP = 0.608, S.E. = 1.475) were high but imprecise, which is not surprising given the relatively small sample size.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Distal and local pQTL for sputum MUC5AC and MUC5B concentration.

Results of association analysis using sputum mucin concentration data from 215 EA SPIROMICS participants are shown. Dashed red line denotes genome-wide significance threshold.

https://doi.org/10.1371/journal.pgen.1010445.g001

For MUC5B, one local pQTL was detected on Chr 11 (rs140324259), and one distal pQTL was located on Chr 4 (rs10001928). One additional MUC5B pQTL (rs6043852), located in the intron of KIF16B on chromosome 20 was detected by testing for the joint effects of SNP + SNP × smoking interactions (joint test p-value = 1.3x10^-9, Fig 2A). Further analysis revealed that the interaction itself contributed substantially to the joint effect (p_interaction = 1.1x10^-7), such that rs6043852 was associated with MUC5B concentration only in subjects that are not current smokers (Fig 2B). Given the relatively low minor allele frequency of rs6043852 (3%), the number of subjects harboring genotypes with the minor allele (A) and of contrasting smoking status was not large (n = 6 A allele carriers in current smokers and n = 6 in the combined never plus former smoker group). Hence this SNP × smoking interaction pQTL must be interpreted with caution.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. A genotype x smoking interaction locus (rs6043852) for sputum MUC5B concentration on Chromosome 20.

A. Locus zoom plot for the genotype x smoking locus (rs6043852). B. MUC5B concentration as a function of both rs6043852 genotype and current smoking status. The not current smoker category includes never smokers and former smokers. Numbers in parentheses on x-axis denote sample size per genotype. Note that while we plot carriers of the minor allele here as one group, the regression model for MUC5B used genotype dosages.

https://doi.org/10.1371/journal.pgen.1010445.g002

The strongest pQTL we identified was for MUC5B on Chr 11. The lead variant, rs140324259, is located approximately 100 kb upstream of MUC5B, in between MUC2 and MUC5AC (Fig 3A). A second variant located in intron 6 of MUC5AC, rs28668859, was also associated with MUC5B concentration; conditional analysis revealed this this signal was partially dependent on linkage disequilibrium (LD, R² = 0.20) with rs140324259 (conditional p-value = 1.6x10^-3). Neither rs140324259 nor rs28668859 are in LD (R² = 0.02 and 0.01, respectively) with the MUC5B promoter variant rs35705950 that is a well-known MUC5B eQTL and is associated with IPF [13]. After adjusting for covariates, rs140324259 genotype explained ∼14% of variation in sputum MUC5B, and each minor allele (C) contributed a ∼2.3 pmol/ml unit decrease in MUC5B (Fig 3B), an effect size that is greater than the effect of current smoking status (yes vs. no, ∼1.4 pmol/ml). Adjusting for disease severity (using GOLD stage) did not materially change these results.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. The chromosome 11 MUC5B pQTL.

A. Regional view of association test results. Four genes not shown due to small size. The lead variant, rs140324259, is approximately 100 kb upstream of MUC5B. B. Effect of rs140324259 genotype on sputum MUC5B concentration. Numbers in parentheses on x-axis denote sample size per genotype. Each C allele yields a 0.8 log (ln) unit decreased in MUC5B, equating to a 2.3 picomol/ml decrease in MUC5B concentration.

https://doi.org/10.1371/journal.pgen.1010445.g003

We asked whether lead MUC5B pQTL, rs140324259, was associated with MUC5B gene expression in SPIROMICS or other studies. In a subset of SPIROMICS participants (n = 144) for whom airway brush RNA-seq data exist [15], we used a tagSNP for rs140324259, namely rs55680540 (LD R² = 0.72 in entire SPIROMICS population), but found no correlation between genotype and MUC5B expression (S8A Fig). No other variants in the region were significantly associated with MUC5B expression (S8B Fig). Additionally, rs140324259 was also not associated with MUC5B expression in the nasal epithelium of subjects with cystic fibrosis [16] or asthma [12], nor was it reported as an eQTL in any tissue in the GTEx dataset [17], including homogenized lung tissue.

Given that power for eQTL detection could be an issue underlying the negative eQTL association results, we asked whether rs140324259 or four variants in LD (rs55680540, rs28668859, rs11604917, and rs76498418) could potentially affect gene expression by altering transcription factor binding or chromatin state using Haploreg [18]. As shown in S2 Table, rs140324259, rs11604917, and rs55680540 are predicted to alter binding of transcription factors, and there is some evidence that rs11604917 and rs55680540 alter chromatin state in relevant cell types or tissues (S3 Table). Perhaps most notably, rs11604917 lies in an enhancer region in multiple cell types and tissues and is predicted to alter binding of the transcription factor RBP-J. The alternate allele (C) of rs11604917 disrupts the consensus sequence at the first position of an almost invariant motif (S9 Fig). Given that RBP-J is part of the Notch signaling pathway that determines ciliated vs. secretory cell fate in murine airways [19], we asked whether rs11604917 was associated with the frequencies of basal, secretory and ciliated epithelial cells, which were estimated using a deconvolution approach on airway brush bulk RNA-seq data from 137 SPIROMICS participants (largely overlapping with the eQTL dataset described above). We applied a previously developed cell type deconvolution method [20] that was shown to perform well on airway samples [21] (see Methods). We did not detect any associations between rs11604917 genotype and frequencies of these cell types, though smoking status was strongly associated with cell type proportions (S4 Table).

Association of MUC5B pQTL with COPD Phenotypes

Subsequently, we tested whether rs140324259 was related to clinically-relevant COPD phenotypes, namely CB and AE. In the subset of SPIROMICS participants with complete phenotype, genotype, sputum MUC5B, and clinical outcomes data (n = 141–143), we found that rs140324259 was not associated with CB at baseline/enrollment (p = 0.25, S5 Table). rs140324259 was not associated with AE in the year prior to enrollment (p = 0.14, Fig 4A) unless we accounted for sputum MUC5B concentration (p = 0.02, Fig 4B and S6 Table). Surprisingly, in this analysis, we found that while MUC5B concentration was positively associated with AE (β₁ = 0.45, p = 0.01, Fig 4B), the effect of rs140324259 genotype (β₂ = 0.74, p = 0.02) was directionally opposite our expectations based on pQTL analysis. That is, the C allele of rs140324259, which was associated with lower MUC5B (γ = -0.77, p = 2.6 x10^-6) and therefore would be expected to confer decreased risk of AE, was associated with increased risk of AE compared to the T allele.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Mediation analysis reveals that rs140324259 exerts effects on exacerbations in the year prior to enrollment through direct and indirect paths with contrasting allele effects.

We leveraged the mediation analysis framework of Baron and Kenny [22] to examine whether rs140324259 exerts effects on exacerbations through MUC5B. Using complete data on 142 subjects, in (A) we tested for the total effect of rs140324259 on acute exacerbations of COPD (“c”). In (B), the mediation analysis framework is shown in which the effect of rs140324259 on acute exacerbations is modeled as the sum of direct (rs140324259 to exacerbations, c’) and indirect paths (rs140324259 to exacerbations via MUC5B (a, b)). Statistical evidence of the indirect path assessed by jointly testing that both rs140324259 → MUC5B (a) and MUC5B → exacerbations (b) are significant using an intersection union test (which is equivalent to testing that γ x β₁ is not equal to 0). β₁ (b) and β₂ (c’) come from the same negative binomial regression model including both rs140324259 and MUC5B as predictors of exacerbations. Note that in this mediation analysis framework, the total effect (c) in part A is the sum of the direct (c’) and indirect paths (a→b) in part B, i.e., c = c’ + (a x b). Thus, because the sign of path a is negative while both b and c’ are positive, the total effect c (in panel A) is necessarily weaker in magnitude.

https://doi.org/10.1371/journal.pgen.1010445.g004

These results suggested the possibility that rs140324259 may exert effects on AE through both direct and indirect paths, the latter via MUC5B (Fig 4B). To examine this further, we employed a mediation analysis approach, based on the framework developed by Baron and Kenny [22], in which the effect of rs140324259 on AE is modeled as the sum of direct (rs140324259 → AE) and indirect paths (rs140324259 → AE via MUC5B). We leveraged an intersection union test [23] to jointly test that both components of the indirect path (rs140324259 → MUC5B and MUC5B → AE) are statistically significant. Indeed, we found evidence that this is the case (p = 0.02), which is consistent with a model of partial mediation by MUC5B. Thus, overall, we conclude from these results that rs140324259 likely affects AE in two ways, both directly and indirectly, but with contrasting allele effects in each case, and overall the net effect is that rs140324259 C allele confers increased risk of AE. We note also that contrasting direction of effects of rs140324259 → MUC5B (-) and MUC5B → AE (+) likely explains why the magnitude of the association between rs140324259 and AE (not accounting for MUC5B) is weak and therefore not statistically significant (Fig 4A) [22].

We then examined associations between rs140324259 and clinical phenotypes in the larger SPIROMICS population for which genotype and clinical data exist but there is not sputum mucin concentration data (n≈1,250). In this sample, rs140324259 was associated with CB at baseline (p = 0.02, Table 2). Similar to results in the smaller subset of subjects described above, the C allele was associated with increased risk of CB (odds ratio (OR) = 1.42; 95% confidence interval (CI): 1.10–1.80). The effect of rs140324259 on AE in the larger SPIROMICS sample with clinical data was examined using both retrospectively and prospectively ascertained data. rs140324259 was not significantly associated with AE in the year prior to enrollment (S7 Table), nor in the year following enrollment (S8 Table). However, in both of these analyses, the results were suggestive and the direction of effect was again positive for the C allele.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Logistic Regression Model of Chronic Bronchitis at Baseline and rs140324259 genotype (n = 1257).

https://doi.org/10.1371/journal.pgen.1010445.t002

We then asked whether rs140324259 genotype was associated with AE over a period of three years of follow up. SPIROMICS participants’ exacerbation frequency was categorized based on a previous study as never (n = 433), inconsistent (n = 331) or consistent (n = 58) over the three years [24] (see Methods). Using a proportional odds model, we analyzed whether rs140324259 genotype distinguished never exacerbators versus inconsistent and consistent AE, and whether rs140324259 genotype distinguished consistent exacerbators versus never and inconsistent exacerbators. We found that rs140324259 genotype was associated with the former contrast (p = 0.03), with the C allele conferring increased risk of being either an inconsistent or consistent exacerbator (S9 and S10 Tables). rs140324259 genotype was not associated with the contrast between consistent exacerbators versus never and inconsistent exacerbators. To simplify the interpretation of the effect of rs140324259 on prospectively ascertained AE, we then dichotomized subjects into two groups, those who experienced AE (inconsistent and consistent) versus those who did not. As shown in Table 3, the rs140324259 C allele conferred increased risk of AE over three years of follow up (OR = 1.41; 95% CI: 1.02–1.94).

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Logistic Regression Model Comparing Exacerbators Versus Non-Exacerbators Based on Prospectively Ascertained Exacerbation Count Over a Three-Year Period (n = 822).

https://doi.org/10.1371/journal.pgen.1010445.t003

Replication analyses

Finally, we analyzed data from the COPDGene study population and UK Biobank in an attempt to replicate results from SPIROMICS. For COPDGene, we utilized phenotype data from COPD cases of European ancestry, and genotypes from whole genome sequencing. Sample sizes in these analyses ranged from 5300–5700 depending on the outcome. In this population of COPD cases, rs140324259 was not associated with CB (OR = 1.08 (95% CI: 0.94–1.24), nor AE (S11–S13 Tables), though we were unable to directly evaluate whether rs140324259 was associated with the exacerbation frequency categories (never, inconsistent, consistent) described in SPIROMICS. In the UK Biobank, however, we found that rs140324259 was associated with two CB-related phenotypes, namely bringing up phlegm/sputum/mucus on most days and cough on most days (Table 4). Importantly, the C allele was enriched in cases vs. controls for these two phenotypes, thus these results are directionally consistent with results from SPIROMICS. Results for other variants in LD with rs140324259 are shown in S14 Table. As UK Biobank results were not adjusted for smoking, we additionally assessed whether rs140324259 was associated with smoking. rs140324259 was either not associated with smoking history variables or was weakly associated, but in these cases the C allele frequency was higher in controls than cases (Table 4), suggesting that the associations between rs140324259 and CB-related phenotypes in the UK Biobank are unlikely to be mediated by, or confounded with, smoking.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Association analysis results for lead MUC5B pQTL variant with CB-related phenotypes and smoking history in the UK Biobank.^*

https://doi.org/10.1371/journal.pgen.1010445.t004

In contrast to the associations we detected between COPD phenotypes and the lead MUC5B pQTL variant, we did not detect any associations with the other MUC5B pQTL (rs10001928), nor the two MUC5AC pQTL (rs75401036 and rs16866419), in SPIROMICS, COPDGene, or the UK Biobank, as shown in S15–S17 Tables.

Ethics statement

Subjects provided informed written consent to participate in the studies described here. Details and institutional review boards for each clinical site are provided in S2 File.

Study subjects and genotype data

The primary analyses presented here are based on study participants in SPIROMICS (ClinicalTrials.gov Identifier: NCT01969344), and a schematic of the SPIROMICS datasets used here is shown in S1 Fig. The study design has been described previously [31]. SPIROMICS participants were genotyped using the Illumina OmniExpress Human Exome Beadchip [32]. Quality controls included testing for sex concordance and removal of SNPs with high genotype missing rates (>5%) and/or violations of Hardy Weinberg equilibrium at p < 1x10^-6. Genotype imputation was performed using the Michigan Imputation Server [33] using haplotypes from Phase3 of the 1000 Genomes Project [34]. Study participants were categorized into either European ancestry (EA, N = 576 or African ancestry (AA, N = 132) groups based on principal components analysis of genotype data and comparisons to 1000Genomes data. We used an adaptive R² threshold to filter imputed variants in each ancestry group based on the minor allele frequency (MAF). For each MAF interval, the R² value was chosen such that the average R² for variants with values larger than the threshold was at least 0.8 (S18 and S19 Tables). We limited our analyses to SNPs with minor allele counts >8, resulting in ∼10 million SNPs in EA and 12 million variants in AA subjects for association with total and specific mucin concentrations.

In COPDGene [35] (ClinicalTrials.gov Identifier: NCT00608764), genotype data for rs140324259 was obtained from whole genome sequencing performed through the TOPMed consortium [36]. Results from the UK Biobank data were obtained from the Pan-UK Biobank analysis (see further description below) [37].

Sputum mucin phenotype data

Sputum mucin concentration: sputum induction and measurement methods have been previously reported [8,9,38]. In brief, hypertonic saline was used to induce sputum, which was then placed in a buffer containing 6 molar guanidine, and stored at 4 degrees. Total sputum mucin concentration was determined using a size exclusion chromatography / differential refractometry measurement approach. For a subset of subjects, MUC5AC and MUC5B concentration was determined using stable isotope labeled mass spectrometry [38]. Data were generated in two batches. In addition to SPIROMICS participants with COPD (n = 439), two additional sets of subjects were also included in the mucin analyses: non-smoking controls (n = 50), and smokers without COPD that are referred to as the “at-risk” group (n = 219). These subjects were included in genetic analysis of sputum mucin concentration but were not included in the analysis of COPD outcomes.

Clinical/phenotype data

We analyzed data on two COPD phenotypes, namely CB and AE, in SPIROMICS. The CB phenotype was ascertained at the first study visit (“baseline”) and was categorized based on participants’ responses to questions regarding frequency of cough and mucus/phlegm production in the St. George’s Respiratory Questionnaire, as described in a previous publication [39]. More specifically, participants were categorized as having CB if they indicated they cough either most days a week or several days a week and bring up phlegm/sputum either most days a week or several days a week. If participants said they only cough or bring up phlegm with respiratory infection or not at all, they were categorized as negative for CB.

The analysis of AE was based on previous work from SPIROMICS [24,31] in which AE were defined as events that required health care utilization (i.e., office visit, hospital admission, or emergency department visit for a respiratory flare-up) involving the use of antibiotics and/or systemic corticosteroids. In COPDGene, the CB phenotype was based on chronic cough and phlegm production for ≥ 3 months/year for 2 consecutive years [5]. For AE, self-reported moderate-to-severe exacerbations in the year prior to enrollment and the number of moderate-to-severe exacerbations ascertained prospectively from longitudinal follow up data were examined. In the Pan-UK Biobank (https://pan.ukbb.broadinstitute.org/), we evaluated results of association analyses for two CB-related phenotypes, namely bringing up phlegm/sputum/mucus daily (yes vs. no, phenocode 22504) and coughing on most days (yes vs. no, phenocode 22502), as well as smoking history variables, which were assessed by questionnaire.

Sputum mucin concentration

Data on total and specific (MUC5AC and MUC5B) mucin concentrations were log-transformed prior to analysis. GWAS analysis was performed using version 0.5.0 of the ProbABEL software [40]. Analysis of total mucin concentration was conducted in each ancestry group separately (N = 576 EA and 132 AA), followed by a pooled analysis of both ancestry groups. In ancestry-specific analyses, main effect SNP models of each mucin phenotype included covariates for the top two principal components of ancestry (PC) obtained from EIGENSTRAT [41], age, sex, batch of mucin quantitation analysis, current smoking status, smoking pack-years, and CB. Results were not materially different when we included up to 10 genotype PCs. For MUC5AC and MUC5B, GWAS was performed in EA subjects only (N = 215) with the same covariates used for total mucin concentration. SNP-based heritability for MUC5AC and MUC5B was estimated on an LD-pruned set of markers from the genotyped data (subsetting to individuals with the relevant phenotype data) using GCTA version 1.92.1. We performed exploratory genome-wide interaction studies of SNP × smoking interactions in which we tested for the joint effects of SNP and SNP × smoking interactions (2 d.f. test) on mucin concentrations in models including the same covariates as above.

eQTL analysis

Airway epithelial gene expression from 144 SPIROMICS participants was analyzed to test whether rs140324259 is an eQTL for MUC5B by performing a genome-wide eQTL mapping as described before in Kasela et al. [15]. Briefly, RNA-seq data from the airway epithelium was normalized, filtered, and transformed using inverse normal transformation. Genotype data was obtained from TOPMed (Freeze 9) [36]. The eQTL regression model for a given gene included sex, four genotype PCs, and 15 PEER factors (probabilistic estimation of expression residuals [42]) as covariates. eQTL mapping was performed using tensorQTL [43] and 10,000 permutations were used to control for multiple testing at false discovery rate (FDR) < 0.05. To look up the eQTL association with MUC5B, we used the proxy SNP rs55680540 because rs140324259 did not pass variant filter quality control.

Cell type deconvolution

Using the same original dataset as was used for eQTL detection in bronchial epithelial brush samples, we performed cell type deconvolution using the non-negative least squares (NNLS) method (R package nnls) [44], which has been shown to perform well in comparison to other deconvolution methods [20]. Application of this method in a previous study of COPD subjects [21] showed good correlations between goblet cell proportions estimated in bulk sequencing data using NNLS and goblet cell numbers based on histological scoring. NNLS bulk deconvolution was carried out using cell type specific signatures derived from bronchial biopsies in six subjects with asthma (in which a larger proportion of goblet cells is identified than in healthy controls [45]). Cell type signature derivation of 14 known cell types was done in the scRNAseq data using AutoGeneS (Automatic gene selection using multi-objective optimization for RNA-seq deconvolution [46]). NNLS was then used to estimate the proportions of these cell types in the bronchial epithelial brush bulk RNA-seq samples of 137 SPIROMICS participants. Cell types included six epithelial cell populations (Ciliated cells, Goblet/Club cells, Submucosal Secretory cells, Basal cells, Basal cycling cells, and Ionocytes), two stromal cell populations (Endothelial cells and Fibroblasts), and six white blood cell populations (Mast cells, Dendritic cells, Alveolar macrophages, T cells, Monocytes, B cells).

We then performed analyses with both rs140324259, the lead MUC5B pQTL, and rs11604917, the variant in LD that is predicted to alter RBP-J transcription factor finding, testing for genotype associations with the frequency of goblet cells, ciliated cells, and basal cells, which are the cell types one would hypothesize should be most relevant to mucin gene expression (we did not test for association with submucosal secretory cells because all values were 0). We used a beta regression model with a data transformation applied to cell type frequencies (y’ = (y(n-1)+0.5)/n) to avoid having values of 0 or 1.

Clinical phenotypes

Chronic Bronchitis (CB): Following the analyses of sputum mucin concentration data, we tested for an association between the lead variant for sputum MUC5B concentration (rs140324259) and CB in the larger SPIROMICS population (N = 1257). Logistic regression models were used for CB, accounting for the top two genotype PCs, age, sex, current smoking status, pack-years of smoking, and FEV1 (% predicted).

Acute Exacerbations (AE): Exacerbation outcomes were modeled using negative binomial regression models including the same covariates as above. Additionally, in the analysis of prospectively ascertained AE, we included AE in the year prior to enrollment as a predictor. Because prior work showed that exacerbation frequency among subjects with COPD in SPIROMICS is not stable [24], we leveraged a previously developed classification system which categorized SPIROMICS participants as never, inconsistent or consistent exacerbators using three years of follow up data [24]. Consistent exacerbators were subjects who experienced at least one acute exacerbation in each of the three years; subjects who had an exacerbation during some but not all of the three years of follow up were defined as inconsistent exacerbators. We analyzed the association between rs140324259 genotype and these three exacerbation groups using a proportional odds model, comparing (1) never exacerbators versus inconsistent and consistent exacerbators, and (2) consistent exacerbators versus never and inconsistent exacerbators. Based on the results of these analyses, we collapsed the exacerbation groups into two categories: ever (combining inconsistent and consistent exacerbators) vs. never exacerbators, then modeled this outcome using logistic regression with covariates for the top two genotype PCs, age, sex, current smoking status, pack-years of smoking, FEV1 (% predicted), and the number of AE in the year prior to enrollment.

Mediation analysis

In the subset of SPIROMICS subjects for which there is complete phenotype data on genotype, sputum MUC5B, and clinical outcomes (N = 141), we tested for evidence of that MUC5B mediates an association between rs140324259 and AE (i.e. rs140324259 → MUC5B → AE), invoking the overall mediation analysis framework of Baron and Kenny [22].

We evaluated a direct path from rs140324259 → AE (c), and an indirect path from rs140324259 → MUC5B (a) and MUC5B → AE (b), while also examining the path from rs140324259 → AE conditional on MUC5B (c’). All regression models included age, sex, two ancestry PCs, current smoking status, pack-years of smoking, and FEV1 (% predicted) as predictors. For (a), we used a linear model for MUC5B in which rs140324259 was coded linearly (0,1,2). For (b), (c), and (c’), we used negative binomial regression models of AE that included rs140324259 (c), MUC5B (b), or both (c’). To formally test for mediation, we leveraged the SNP mediation intersection-union test (SMUT) [23] which jointly tests for non-zero parameter estimates from models for (a) and (b), which is equivalent to testing that a x b is not equal to 0 in the Baron and Kenny framework.

Disclosure

The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Heart, Lung, and Blood Institute or the National Institutes of Health.