Introduction

Measures of pulmonary function provide important clinical and research tools for evaluating lung disease and other morbidities1. Pulmonary function is used to diagnose chronic obstructive pulmonary disease and to monitor the severity and progression of many lung conditions. Furthermore, even within the normal range, pulmonary function is a risk factor for mortality and morbidity from various conditions2,3,4,5. Genome-wide association studies (GWASs) have identified many single nucleotide variants (SNVs) for pulmonary function after adjusting for cigarette smoking, a well-established risk factor for reduced pulmonary function, as a confounder and potential effect modifier6,7,8,9,10,11,12,13,14. The interplay between genetic and environmental factors likely plays a role in complex traits, including chronic lung diseases. A large GWAS meta-analysis has shown the existence of SNV-by-smoking interaction in relation to two measures of pulmonary function (forced expiratory volume in one second, FEV1, and the ratio of FEV1 to the forced vital capacity, FEV1/FVC) by performing joint analyses of genetic main effects and interaction effects of single SNVs and their interaction with smoking14. However, there is a lack of gene-by-environment (GxE) interaction analysis involving low-frequency (minor allele frequency [MAF] 1% to 5%) or rare variants (MAF less than 1%)15, which are typically not well captured or imputed by previous GWAS meta-analyses of pulmonary function. The more recent use of genotyping platforms including lower frequency and rare variants enables better assessment of the role of rare genetic variation in disease.

Due to the low power to detect association with rare variants, a commonly adopted strategy in analyzing rare variant-phenotype associations is to group variants according to genes or genomic regions and test whether the grouped variants are associated with the phenotype16. While single-variant-based interaction tests for common variants are well-established17,18, methods for detecting rare variant GxE interactions are relatively new. Recently developed novel approaches for testing rare variant GxE interaction effects include a joint test that allows for simultaneous testing of the genetic main and interaction effects as well as the ability to assess gene-based GxE interactions only for both related and unrelated individuals19,20,21,22. The joint test and the test for the interaction terms only are both important for different scientific hypotheses. The joint test assesses genetic associations with pulmonary function, accounting for heterogeneity of genetic effects in individuals with different environmental exposures (in our context, smoking behaviors). The test of the interaction terms only can identify genes that modify the effect of smoking on pulmonary function. For single variant analysis, an interaction requires at least four fold larger sample size than a genetic main effect of comparable magnitude23. Similarly, for gene-based interaction tests, larger sample sizes are required for detection, especially for rare variants. This study is the first to incorporate GxE interactions in modeling rare and low-frequency genetic variants and smoking effects on pulmonary function based on large samples. In discovery, we used genetic and phenotypic information from the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE)24 and SpiroMeta Consortia, and we sought replication of findings in the UK BiLEVE and UK Biobank studies.

Results

Characteristics and sample sizes of the study cohorts are presented in Table 1. We focused on two smoking categories throughout: ever-smokers vs never-smokers and current smokers vs never-smokers and past-smokers combined. The percentage of ever-smokers per study ranged from 30 to 61% in 44,429 European Ancestry participants and the percentage of current smokers per study ranged from 6 to 25%.

Table 1 Descriptive statistics for each participating cohort.
Full size table

Quantile–quantile plots of the meta-analysis results for the interaction and joint tests from the six combinations of smoking variables status (ever smoking or current smoking) and pulmonary function measures (FEV1, FVC, and FEV1/FVC ratio) of discovery stage are presented in Figure S1. Some quantile–quantile plots suggest inflation in the gene-by-current smoking interaction analyses, which could be due to the low frequency of current smokers. The few studies with larger inflation factors (λ > 1.5) in testing interactions between current smoking status and a specific lung function measure were not included in the corresponding meta-analysis of this smoking-trait combination (Table S1).

In the discovery-stage meta-analyses, GPR126 was significantly associated with FEV1/FVC in the gene-by-current-smoking (p-value = 1.9E−09) and gene-by-ever-smoking analyses (p-value = 2.8E−08) based on joint tests. In both replication studies, GPR126 was also statistically significant in the joint tests for gene-by-current-smoking (p-value = 1.1E−16 in UK BiLEVE, 1.4E−30 in UK Biobank) and gene-by-ever-smoking (p-value = 1.7E−16 in UK BiLEVE, 5.2E−29 in UK Biobank, Table 2). The average number of rare variants in GPR126 across studies was 18.8 (range from 5 to 25). The p-value of the interaction test was not significant for either smoking exposure evaluated (p-value > 0.5), suggesting that these joint test results were largely due to the genetic main effects.

Table 2 Significant genes in the meta-analyses (significant threshold = 5.7E−07).
Full size table

A common intronic variant rs3817928 (MAF = 22%) in GPR126 was previously found to be significantly associated with FEV1/FVC among European ancestry25. Conditioning on this common variant rs3817928, the joint test of gene-by-current-smoking analysis for GPR126 remained nominally significant in the discovery (p-value = 2.4E−03) and genome-wide significant in the replication studies (p-value = 3.9E−09 in UK BiLEVE and p-value = 9.6E−17 in UK Biobank, Table 3). The joint test of gene-by-ever-smoking analysis conditioning on rs3817928 was also nominally significant in the discovery stage and genome-wide significant in the replication stage (p-value = 0.01 in discovery stage, p-value = 4.3E−09 in UK BiLEVE, and p-value = 2.5E−16 in UK Biobank, Table 3).

Table 3 Conditional analysis of GPR126 (adjusting for a common variant rs3817928).
Full size table

To further determine whether any single low-frequency or rare variant may be driving the aggregate variant test results, we conducted single variant interaction only and joint test analyses for all the SNVs in GPR126 and report the most significant SNV, rs17280293 (MAF = 3%), in Table 4. Figures S2 and S3 show the regional association plots for GPR126. The significant joint test results for rs17280293 were replicated in the UK BiLEVE and UK Biobank studies (Table 4). This SNV is predicted to be deleterious by SIFT score (score = 0.001) in dbNSFP26.

Table 4 Single variant analysis for the identified significant genes.
Full size table

SNV rs17280293 is in low linkage disequilibrium (LD) with the previously reported common variant rs3817928 in GPR126 (R2 = 0.13 based on 1000 Genomes database Phase 327), thus rs17280293 remained highly significant in the single-variant-based analysis when conditioning on rs3817928 (Table 3). Recently, it was shown that the rs17280293 was the leading signal that drove a gene-based association between GPR126 and FEV1/FVC when studying genetic main effects15. It was also found to be the most significant signal in the GPR126 region associated with diffusing capacity of the lung28. Our analyses on the interaction-only and joint analyses further highlighted the importance of this low-frequency variant.

PDE3B was significantly associated with FEV1 in the interaction-only test (p-value = 1.9E−06) and joint test (p-value = 2.9E−07) in the gene-by-current-smoking analyses (Table 2). The p-value of the interaction test was highly significant, suggesting that the joint test result was at least partially driven by the interaction effects. However, the interaction effect was not replicated in either UK BiLEVE or UK Biobank. Single-variant-based interaction and joint analyses were performed on the variants within PDE3B (see the regional association plots in Figures S4 and S5). The top two variants are present in Table 4, among which variant rs142935352 (MAF = 0.6%) in PDE3B had a significant SNV-by-current smoking interaction in relation to FEV1 in the discovery studies and was nominally significant in the UK BiLEVE replication study for both the interaction-only (p-value = 0.01) and joint tests (p-value = 0.02). The other top variant rs61736639 (MAF = 0.7%) was not significant in the replication datasets. According to the Genotype-Tissue Expression (GTEx) project (29, accessed on 07/18/2021) both GPR126 (median transcripts per million, TPM = 20.2) and PDE3B (median TPM = 8.8) were expressed in lung tissue, relative to a minimal expression threshold of TPM > 0.130.

Discussion

Most genome-wide analyses of pulmonary function have focused on common variants or genetic main effects25,31. However, the gene-level association could be missed if the signal is driven by rare variants or there is interaction with environmental exposures. Therefore, we focused on rare and low-frequency exonic genetic variants and their interaction with cigarette smoking, the major environmental risk factor for reduced lung function. Specifically, we performed gene-based tests of GxE interaction only and joint tests assessing genetic main effects and interaction effects. The joint analysis has been shown to be nearly optimal across a variety of models for common variants17. In some scenarios, the GxE interaction only tests could have higher statistical power than the joint tests19,21. To boost the statistical power for discovery, we combined data across thirteen studies with 44,429 European ancestry individuals through meta-analyses. Significant genes were examined in two replication studies with a total of 104,611 European ancestry individuals. This study is the first and the largest to incorporate GxE interactions in modeling rare and low-frequency variant genetic and smoking effects on pulmonary function. However, statistical power would continue to be improved with increased sample size. Table S3 and S4 provide the top 5 genes for the interaction only and joint tests, which may include the putative genes that did not reach genome-wide significance in the study.

Our results for the GxE joint analysis support the involvement of rare variants in a locus, GPR126. We identified one low-frequency variant rs17280293 which was successfully replicated in our study and was previously reported by Jackson et al15 in a genetic main effect analysis. The SNV is in low LD with the known common variant, suggesting a distinct role of the rare variants. Neither gene-based nor single variant tests identified interaction with smoking for this SNV or this gene. Additionally, we found that PDE3B was significantly associated with FEV1 in the GxE interaction and joint association analysis in the discovery studies, however neither the interaction or joint effect gene-based tests replicated, although one leading rare variant rs142935352 reached nominal significance in UK BiLEVE for the interaction and joint tests. PDE3B had not been reported to be associated with lung function but is believed to be an important regulatory factor in modulating the inflammatory response and expressed in the lung32. One possible explanation is that smoking increases the inflammatory response in the lung, leading to the gene-by-smoking interaction. The signal of PDE3B in the discovery dataset was mainly driven by two SNVs (rs142935352 and rs61736639), both quite rare (MAF < 0.5% in EA), thus difficult to replicate even with large samples. We found that it was challenging to replicate rare signals in general, as rare variants varied across different studies. A variant present in one study was likely not included in the other study (for example, being monomorphic or having high missing rates).

Our study has limitations. First, the results were obtained from European ancestry populations, although we recognized the importance of conducting multi-ancestry studies to identify novel association signals31. Second, a few quantile–quantile plots suggested inflation in the gene-by-current-smoking interaction and joint test statistics (Figure S1). GxE analysis is well-known to be prone to systematic inflation for common variants and there could be various factors driving the inflation33,34,35. It is not surprising to see such inflation in rare-variant-based GxE analysis. We speculate that the inflation occurred in some studies due to a combination of a low proportion of current smokers and relatively small sample size, where the asymptotic property of the test statistics may not hold. For example, LBC1936 with a sample size of 970 and 11% of current smokers had an inflation factor > 1.5 for gene-by-current smoking interaction tests but not gene-by-ever smoking interaction (Table S1). Although we took a number of steps to reduce the inflation (adjusting for principal components, excluding studies with large inflation factors, excluding phenotypic outliers followed by inverse normal transformation of the spirometric measures of pulmonary function, excluding a study for a particular gene if it had low minor allele count (< 20), and using dichotomized smoking risk factors21), there could still be some inflation. Third, we examined exonic rare variants, but we note that signals from rare variants in the regulatory regions are likely to interact with smoking. Inclusion of rare regulatory variants could improve our understanding of disease mechanisms and warrants further investigation. Fourth, our analysis did not take into account the type of cigarette smoking, which could increase the difficulty of replication due to the potential heterogeneity of GxE interaction effects. Fifth, rare and low-frequency variants in the two replication datasets UK BiLEVE and Biobank Study were imputed using the 1000 Genomes and UK10K haplotype reference panel36. Although the imputation quality had been examined through a pseudo-GWAS simulation36, it may still be suboptimal comparing with sequencing data, thus increasing the replication difficulty. In our case, SNVs included in GPR126 and PDE3B all had INFO score > 0.8 and the leading SNVs had INFO score > 0.95 (Table 4), thus less likely to be affected by the genotyping array platform.

The strengths of the study include the relatively large sample size with genotyping of rare and low-frequency variants. In addition, we employed recently developed methods for incorporating environmental interactions in the study of the influence of rare variants in disease to decrease the multiple testing burden, potentially increase the statistical power. We found that one of the first discovered genes associated with pulmonary function in GWAS of common variants, GPR126, likely has low-frequency variants contributing to its role in this phenotype and other putative genes. The gene was successfully replicated but the signal may be driven by the genetic main effect.

Materials and methods

Ethics statement

All the studies contributed to our analyses have their protocols approved by the respective local Institutional Review Board with the details in the Supplementary Materials. All participants wrote informed consent for the genetic studies and all experiments were performed in accordance with relevant guidelines and regulations.

Cohort studies

We combined data from 13 studies for the discovery stage, either from the CHARGE Consortium or the SpiroMeta Consortium: the Age, Environment, Susceptibility (AGES) study, Atherosclerosis Risk in Communities (ARIC) study37, British 1958 Birth Cohort (1958BC)38, Cardiovascular Health Study (CHS)39, Framingham Heart Study (FHS)40,41, Health, Aging, and Body Composition (HABC) study42, Northern Finland Birth Cohort of 1966 (NFBC1966)43, Multi-Ethnic Study of Atherosclerosis (MESA)44, Rotterdam Study (RS)45, Study of Health in Pomerania (SHIP)46, the Netherlands Epidemiology of Obesity Study (NEO)47, Lothian in Birth Cohort 1936 (LBC1936) study48, and Cooperative Health Research in the Region of Augsburg (KORA) study49. The discovery studies consisted of 44,429 individuals. Replication was conducted in two studies: UK BiLEVE Study (n = 45,133) and UK Biobank (n = 59,478)50. Study descriptions are provided in the Supplementary Note of Jackson et al., 201815 and Artigas et al., 201551.

Pulmonary function measurements and smoking information

Spirometric measures included FEV1, FVC, and their ratio (FEV1/FVC). Details of the spirometry have been previously reported for the discovery15 and replication studies52. Because outlier phenotype values can be influential in rare variant analyses, pulmonary function values with residuals after regressing out covariates (listed below) on each lung function trait larger than four standard deviations from the mean were examined by investigators from each cohort expert in the collection of pulmonary function data. Data points were evaluated regarding whether they were biologically plausible for that individual as opposed to more consistent with data errors. To ensure the normality of the FEV1, FVC, and FEV1/FVC, a ranked-based inverse-normal transformation was performed on each pulmonary function trait by the individual participating studies. For FEV1/FVC, the ratio was first calculated prior to transformation.

Cigarette smoking can have important adverse effects on pulmonary function. Smoking exposure was ascertained by questionnaire at the same time as the pulmonary function tests used in the analyses. Study participants were classified in two different ways: ever smokers vs never-smokers and current smokers vs never-smokers and past-smokers combined. Current smokers were defined as individuals who smoked at least one cigarette per day within the prior 12 months, past smokers were defined as smoking at least one cigarette per day but had stopped at least 12 months, and never smokers reported never having smoked. Ever smokers were a combination of current and past smokers.

Genotyping and quality control

Genotyping of the discovery stage studies was performed mainly using the Illumina HumanExome BeadChip (Table S2). To improve accurate calling of rare variants, genotyped data from nine studies (ARIC, FHS, RS, CHS, AGES, LBC1936, and MESA) were called using GenCall in Illumina’s Genome Studio and the curated clustering files from the CHARGE joint calling effort53. The remaining studies (1958BC, KORA, and SHIP) called their genotypes in accordance to the UK exome chip consortium best practices36 using zCall54.

All studies performed comparable sample-level quality control steps and removed individuals according to the standard quality control metrics, i.e., sex mismatch, duplicate pairs, unexpected relatives, missing pulmonary function measurement, or missing covariate. Other than the standard variant-level quality control step53, we retained only nonsynonymous variants annotated by dbNSFP26 with MAF less than 5% among EAs and excluded monomorphic variants, variants with missing rates larger than 5%, and variants on the sex chromosomes. These low-frequency and rare variants were grouped by gene region in the aggregate variant tests26. Missing genotypes were imputed using a random draw from the binomial distribution with two trials and success probability equal to the estimated MAF. Genes with one or zero variants or overall cumulative minor allele count less than 20 were excluded.

The replication samples were genotyped using the Affymetrix UK BiLEVE or UK Biobank arrays, which both include substantial overlap with the Illumina Human Exome BeadChip55. Thorough sample and genotype quality control steps were undertaken before imputation to a combined 1000 Genomes56 and UK10K Project reference panel57. Following imputation, SNVs were excluded if they had imputation INFO score ≤ 0.5 or minor allele count < 3. Full details of the quality control and imputation procedure of the UK BiLEVE/UK Biobank genotype data were described elsewhere52.

Statistical analysis

Since single-variant tests are usually underpowered for analyzing rare genetic variants, we used an aggregate variant test that analyzes multiple rare variants in a gene. For studies other than FHS, we utilized the R package rareGE58 to conduct gene-based interaction and joint tests that simultaneously test the genetic main effects as well as GxE interaction effects. As a set-based variance component test, rareGE has been shown to be powerful across different underlying GxE association patterns19. For FHS, we used a modified rareGE joint test which incorporates correlation among family members by including a random intercept with covariance structure proportional to the kinship matrix19,59. The rareGE interaction, joint and modified joint tests are briefly described in the Supplementary Materials. Both smoking variables were included in the linear regression model as covariates, while one of them was modeled as the primary exposure variable of interest and was tested for interactions. This resulted in a total of six pairs of pulmonary function traits (transformed FEV1, FVC, and ratio) and smoking status (ever vs never smoking or current vs past plus never smoking). Additional covariates included age, age-squared, sex, height, height-squared, recruitment sites, and the first 10 principal components accounting for any underlying population substructure. Principal components were constructed by each study using GWAS SNVs or exome-chip ancestry informative markers if GWAS not available55. Weight was only included as an adjustment variable for FVC as it is more strongly influenced by adiposity. Study-specific results were meta-analyzed by combining the p-values of the gene-based tests from each discovery study using Stouffer’s weighted Z-score method60: (1) transform p-values to z-values by the inverse CDF of a Gaussian distribution, i.e., \(Z_{i} = {\Phi }^{ - 1} \left( {P_{i} } \right)\), (2) weight the z-values of each study by the sample size, i.e., \(Z = \frac{{\mathop \sum \nolimits_{i = 1}^{k} w_{i} Z_{i} }}{{\sqrt {\mathop \sum \nolimits_{i = 1}^{k} w_{i}^{2} } }}\), where \(w_{i} = \sqrt {N_{i} }\) and \(N_{i}\) was the ith study’s sample size. The final meta-analysis p-value was calculated as \(P = \phi \left( Z \right)\). Studies with genomic control inflation factor λ > 1.5 were excluded.

An a priori significance threshold was defined using the Bonferroni correction for the number of tests evaluated. Specifically, we used a significant threshold at 5.7E−07 based on 14,591 genes for 6 smoking status-lung outcome combinations. Significant genes were evaluated for replication in the UK BiLEVE and UK Biobank studies. To identify specific low-frequency or rare variants driving the signals from the aggregate variant tests, we conducted single variant analyses for the significant genes in both the discovery and replication datasets. Additionally, conditional analyses were performed for significant replicated genes (GPR126 herein) with the previously reported leading common SNV, where the reported signal drivers were added in the gene-based test as covariates. For conditional analyses, p-values less than 0.05 were considered as nominally significant. We further investigated the tissue-specific gene expression of the significant genes on the GTEx Project portal (https://gtexportal.org/).