Targeted high-throughput sequencing of candidate genes for chronic obstructive pulmonary disease

The OLIN studies are an on-going research program focused on asthma, allergy and COPD. It started 30 years ago [9] and now involves more than 50,000 subjects from northern Sweden. Within OLIN, a COPD-cohort was identified at re-examination of several cohorts in 2002–2004 [10]. At recruitment, COPD (n = 993) was defined using the fixed ratio of FEV₁ / FVC < 0.70 (forced expiratory volume in 1 s / forced vital capacity). When calculating the ratio FEV₁ / FVC, the highest values of FEV₁ and the highest value of forced vital capacity (FVC) or slow vital capacity (SVC) were used. This has support in the GOLD documents [1] and is acknowledged in the recent ERS task force guidelines for epidemiological studies on COPD [11]. An age and gender matched control population (n = 993) without obstructive lung function impairment was also recruited [10]. Since 2005 the OLIN COPD cohort with corresponding controls is followed up annually with a basic program including spirometry and interviews regarding symptoms and morbidity [12]. We initially selected 96 COPD cases (18 non-smokers, 43 former smokers and 35 smokers) from those who had an FEV₁ < 80 % of predicted value in 2005 and either FEV₁ / FVC < LLN (lower limit of normal) in 2010 or were rapid decliners with an annual FEV₁ decline of ≥ 60 ml between 2005 and 2010. We also identified a set of 96 age- and gender-matched controls (33 smokers and 63 former smokers) with normal lung function. These 96 cases and 96 controls are henceforth termed the OLIN discovery sample (Table 1). Furthermore, we defined an OLIN replication sample consisting of individuals from the OLIN COPD study for which DNA was available (n = 1,534). From this group we classified individuals as cases when FEV₁ / FVC was lower than LLN in 2010, or if they had a yearly FEV₁ decline from 2005 to 2010 of at least 60 ml (n = 256). The remaining individuals were used as a reference group (n = 1,278). The physiological parameters for the OLIN replication sample included average age (cases 64 years, SD 11; controls 66 years, SD 11; P = 0.0012), gender (cases 123 females, 133 males; controls 569 females, 709 males; P = 0.30), smoking habits (cases 11 pack/year, SD 15; controls 23 pack/year, SD 17; P = 4.2 × 10⁻⁹), weight (cases 73 kg, SD 15; controls 77 kg, SD 14; P = 1.2 × 10⁻⁴) and height (cases 168 cm, SD 9; controls 168 cm, SD 10; P = 0.8). The phenotype description of this sample included measures of FVC (cases 3.29, SD 1.01; controls 3.50, SD 1. 03; P = 0.0031), FEV₁ (cases 1.90, SD 0.67; controls 2.64, SD 0.81; P = 9.5 × 10⁻⁴³) and FEV₁% of predicted values (cases 67 %, SD 17 %; controls 95 % SD 16 %; P = 5.3 × 10⁻⁷³), as well as the FEV₁/FVC ratio (cases 0.57, SD 0.08; controls 0.75, SD 0.07; P = 2.0 × 10⁻¹⁰⁸) when investigated the fifth year of the evaluations. For description of individual reference values, see Additional file 1.

In total 22 genes implicated in lung development, 61 genes and 10 genomic regions previously associated with COPD (Additional files 1 and 2), were investigated using targeted sequencing of captured genomic regions (HaloPlex Protocol Version A, Agilent Technologies, Santa Clara, CA). Regions of 1.5 kb of genomic sequence, including specific intergenic polymorphisms, was also included in the design. The regions of interest (ROI) were designed to target all known exons of major/known transcripts and at least 20 base pairs (bp) of intronic sequences flanking each exon. The sequence capture design included 953 target regions spanning 204.384 bp with 95.9 % (196.066 bp) coverage an average. Captured genomic regions were subjected to high throughput paired-end (100 bp read module) sequencing (HiSeq2000, Illumina, San Diego, CA) at the Science for Life Laboratory in Uppsala, Sweden. Sequence reads were aligned to the hg19 reference genome and single nucleotide variants (SNVs) were called using GATK Unified Genotyper (GATK bundle v.2.2) [13]. Next, we enriched for high quality SNVs by removing SNVs with low confidence (QD < 1.5), Phred scaled quality score (<50) and SNVs within SNV clusters. These high quality variants are henceforth referred to as ‘variants’. We also removed individual cases and controls with sequencing read depth consistently < 10 reads. The strategy for filtering and quality controls are illustrated in Fig. 1.

Test for genetic association and genetic effect was performed for each predicted variant separately using the discovery sample. In addition, rs8040868 and rs11728716 were tested for association using the available OLIN sample (replication sample) as, according to RegulomeDB, these two markers present with potential functional effects on gene regulation (Table 2). Tests for allelic association of individual variants with COPD were performed using the Fisher’s exact test. Results were considered statistically significant when p < 0.05. No adjustment for multiple testing was performed in these analyses. Effect size was measured using odds-ratios (OR) with 95 % confidence intervals (CI).

Visualization SNPs associated with COPD located in the CHRNA3/5 region was made using LocusZoom v1.1 [14] available on http://​locuszoom.​sph.​umich.​edu/​locuszoom/​. RefSeq gene/transcript case–control tests for aggregation of genetic variants in the targeted genomic regions were performed using PLINK/SEQ v0.1 [15]. Tests were divided into an analysis of rare variants with minor allele frequency (MAF) < 5 % or common variants (MAF ≥ 5 %). The UNIQ test, which identifies unique risk alleles, was utilized using default parameters to count the total number of alleles found only in cases (risk variants). Similarly, the SKAT burden tests, which assesses excess of rare alleles in cases compared to controls, was also utilized. Since both UNIQ and SKAT burden are 1-sided tests, we also swapped the phenotype information and analyzed the effects in both directions (excess of alleles in cases or controls) separately to capture evidence of both risk and protective alleles. Due to the matched design of the case and control groups no covariate adjustments (age, sex, pack-years) were performed in analysis using the discovery sample.

Linkage disequilibrium information of associated variants was extracted from genotypes from the sequencing analysis using Haploview 4.2 [16].

To study the possible effect of associated variants on gene expression, we used information from RegulomeDB, a database that combines ENCODE data sets (chromatin immunoprecipitation sequencing (ChIP-seq) peaks, DNase I hypersensitivity peaks, DNase I footprints) with additional data sources (ChIP-seq data from the NCBI Sequence Read Archive, conserved motifs, expression quantitative trait loci (eQTL), and experimentally validated functional variants) [17]. A scoring system is based on the confidence of the functionality of variants, a lower score corresponding to stronger confidence. Subcategories are used to denote additional functional annotations. Combined Annotation Dependent Depletion (CADD) scores were used to assess potential structural and functional effect of associated nonsynonymous variants [18].

The existence of expression quantitative trait loci (eQTLs) was investigated as previously described using genotyping and gene expression data from 1,111 patients who underwent lung surgery at one of three sites, Laval University (discovery sample), University of British Columbia, and University of Groningen (replication sample sets) (referred to as Laval, UBC, and Groningen) [19, 20]. The eQTL data is derived from non-tumour lung parenchymal samples and expression data were adjusted for age, gender, and smoking status. Estimated P-values for each region were Bonferroni-corrected for multiple testing based on the number of SNPs and probe sets (number of SNPs x number of probe sets) and were considered significant if corrected p < 0.05.

The characteristics of each sample are listed in this section as value ± standard deviation. Cases and controls were matched for age (cases: 68 ± 10 years; controls: 66 ± 11 years; p = 0.15), gender (cases: 35 females, 61 males; controls: 31 females, 65 males; p = 0.65) and smoking habits (cases: 26 ± 19 pack/year; controls: 28 ± 12 pack/year; p = 0.53). Both groups were also closely matched for weight (cases: 76 ± 15 kg; controls: 77 ± 15 kg; p = 0.85) and height (cases: 169 ± 9 cm; controls: 169 ± 9 cm; p = 0.7). No non-smokers were included in the control group to avoid false negative results. The cases presented a significant reduction in lung function consistent with moderate or severe COPD. This is illustrated by a reduced FVC (cases: 2.69 ± 0.95 L; controls: 3.96 ± 0.90 L; p = 6.3 × 10⁻¹⁸), FEV₁ (cases: 1.50 ± 0.59 L; controls: 3.11 ± 0.69 L; p = 4.0 × 10⁻⁴¹) and FEV₁% of predicted values (cases: 53 ± 14 %; controls: 110 ± 9 %; p = 2.5 × 10⁻⁸¹), as well as the FEV₁/FVC ratio (cases: 0.56 ± 0.11; controls: 0.79 ± 0.05; p = 2.5 × 10⁻⁴⁵) when investigated the fifth year of the evaluations (Table 1).

We identified 2,151 SNVs after analysis of the sequenced target regions. After variant and sample quality control procedures, 1588 SNVs and 68 cases and 66 controls were retained in the downstream analysis (Fig. 1). Out of the 1588 variants, we identified 37 variants with significantly different allele frequencies in cases and controls (henceforth referred to as ‘associated variants’) (Table 2). We initially detected two novel variants in the discovery sample: GRCh37.p13, 5:g.157002804C > G in the ADAM19 gene and GRCh37.p13, 7:g.73477874C > A in ELN. However, using Sanger sequencing of the same sample we excluded both variants, as they were monomorphic.

Three of the associated variants were shown to be unique to controls including missense variant rs61741262 (p.Asn1722Ser) in TNS1. The most significantly associated variants were all intronic (GSTCD, rs11728716, p = 6.5 × 10⁻⁵, OR = 7.4 (2.5-22.5) and MMP12, rs632009, p = 6.7 × 10⁻⁴, OR = 2.5 (1.5-4.1).

Although the majority of the associated variants were intronic (or intergenic), five were protein-coding (Table 2). Of these, two variants predicted amino-acid substitutions (missense variants): p.(Thr100Met) in MACROD2 and p.(Asn1722Ser) in TNS1 respectively. The p.(Asn1722Ser) variant could be potential damaging based on a relative high CADD score or 12.49. Of the coding variants, we found that rs2143390, predicting a p.(D373=) in GPR126 (p = 0.005, OR = 7.9 (1.6-37.6)), rs6413520 in SFTPD (p.(Ser45=), p = 0.036, OR = 8.2 (1.0-66.4)), rs8040868 in CHRNA3 (p.(Val53=), p = 8.8 × 10⁻³, OR = 2.0 (1.2-3.2)) and rs41275442 in MACROD2 (p.(Thr100Met), p = 0.049, OR = 2.1 (1.0-4.5)) conferred moderate to high risk for COPD (Table 2).

We tested the presence of variants uniquely found in cases or controls as well as gene burden tests in cases against controls as specified in PLINK/SEQ v0.1 using standard settings. We noted that the ADAM19, WNT2, CHRNA5, NOS3 and PTCH1 genes all harbor rare variants (MAF < 5 %) uniquely found in cases (Additional file 3). Conversely, the FGF8, CTNNB1 and HHIP genes contain rare variants uniquely found in the control sample (Additional file 3). Neither gene burden analysis (SKAT) or analysis of rare alleles (MAF < 5 %) yielded significant results. However, by performing a joint analysis with only common alleles (MAF ≥ 5 %) in target regions using SKAT, we showed a significant gene burden for the genes GSTCD, FGF1, ELN and ESR1 (Additional file 4).

We identified five regions with associated variants in pairwise LD (r² > 0.7, D´ = 1.0). The regions were located at the GSTM1 gene locus on chromosome 1 (rs72989301-rs111436983), GSTCD on chromosome 4 (rs72671840-rs72671858), PDE4D on chromosome 5 (rs3805557- rs3805556- rs1553114), MTHFDIL on chromosome 6 (rs803451- rs803448) and TRPV4 on chromosome 12 (rs59870578-rs59940634) (Additional file 5). Furthermore, the variant rs8040868 on chromosome 15q21.1 is in pairwise LD (r² = 0.76, D’ = 1.0) with rs16969968, a nonsynonymous variant previously associated with expression of the CHRNA5 gene [21]. The rs16969968 variant was included in our capture design but it did not reach significant association in the OLIN discovery sample (OR 1.6; p = 0.07) (Additional file 6).

According to RegulomeDB, all 37 associated variants were located within known and predicted regulatory elements in intergenic regions (Table 2). We noted that a variant in CHRNA3 (rs8040868) and a variant in GSTCD (rs11728716) each showed a RegulomeDB score of “1f”, denoting the presence of transcription binding site or DNAse peak.

According to RegulomeDB, the rs8040868 (CHRNA3) and rs11728716 (GSTCD) variants could present with potential functional effects on gene regulation. To determine if these variant could represent eQTL, we analysed the genotypes and gene expression data in the discovery sample (Laval) as well as replication samples (UBC and Groningen). One of these variants, rs8040868:C > T, was confirmed to be significantly associated with gene expression of the nearby gene CHRNA5 in all three data sets, with the C allele (minor allele) associated with lower CHRNA5 expression (Fig. 2 and Additional file 7). Interestingly, we could also see a high correlation between rs8040868 and expression of an anti-sense transcript (AF147302) of unknown function from the adjacent IREB2 gene region (data not shown). AF147302 is likely a result of strong bi-directional promoter activity in this region [22].

We also investigated pulmonary data from the replication sample (n = 1,534; cases = 256, controls = 1278). Analysis using RegulomeDB predicted both rs11728716 and rs8040868 variants as being functional (score 1f for both variants). We therefore selected these two variants for genotyping in all available OLIN samples (n = 1,534). The frequency of the rs8040868-C allele was 35 % in the reference group (n = 1,278) and 42 % in the cases (n = 256) resulting in a significant association (p = 0.003) and an OR of 1.4 (95 % CI 1.1-1.7) for COPD. SweGen variant frequency database reports a 39 % frequency of rs8040868 in 1000 whole genomes representing a cross-section of the Swedish population (https://​swefreq.​nbis.​se). The frequency of the homozygous rs8040868-CC genotype was 12 % in the reference group and significantly higher (18 %) in the cases (p = 0.018). When comparing smoking status using all genotyped individuals, no significant difference in allele frequency between neither the groups ‘non-smokers’ (n = 589) and ‘ever smokers’ (n = 943) (OR 1.1 95 % CI 1.0-1.7; p = 0.09) nor between the groups ‘current smokers’ (n = 312) and ‘former smokers’ (n = 631) (OR 1.2 95 % CI 1.0-1.4; p = 0.11) was seen. The latter test was used to assess nicotine dependency and aptitude for smoking cessation under the assumption that a genetic variant associated with these traits would be underrepresented in a former smoking group as compared to a group of current smokers, i.e., harder to quit smoking. The tests for association with smoking must however be taken with caution as the confidence intervals are wide and a larger sample size would be needed for replication.

Analysis of rs11728716 using the OLIN replication sample (n = 1,534) revealed no association with COPD (p = 0.07; OR = 2.2). In order to test if rs11728716 is associated with severe COPD, we stratified the available COPD cases based on severity and selected cases with FEV1%pred < 40 % and FEV1/FVC < LLN. Our results show a significant association (p = 0.017) between rs11728716 and the group of severe COPD (n = 14). The allele frequency of rs11728716-A was 10 % among cases with severe COPD and 4 % in the controls.