|Home | About | Journals | Submit | Contact Us | Français|
We tested whether genetic factors distinctly contribute to either development of coronary atherosclerosis or, specifically, to myocardial infarction in existing coronary atherosclerosis.
We did two genome-wide association studies (GWAS) with coronary angiographic phenotyping in participants of European ancestry. To identify loci that predispose to angiographic coronary artery disease (CAD), we compared individuals who had this disorder (n=12 393) with those who did not (controls, n=7383). To identify loci that predispose to myocardial infarction, we compared patients who had angiographic CAD and myocardial infarction (n=5783) with those who had angiographic CAD but no myocardial infarction (n=3644).
In the comparison of patients with angiographic CAD versus controls, we identified a novel locus, ADAMTS7 (p=4·98×10−13). In the comparison of patients with angiographic CAD who had myocardial infarction versus those with angiographic CAD but no myocardial infarction, we identified a novel association at the ABO locus (p=7·62×10−9). The ABO association was attributable to the glycotransferase-deficient enzyme that encodes the ABO blood group O phenotype previously proposed to protect against myocardial infarction.
Our findings indicate that specific genetic predispositions promote the development of coronary atherosclerosis whereas others lead to myocardial infarction in the presence of coronary atherosclerosis. The relation to specific CAD phenotypes might modify how novel loci are applied in personalised risk assessment and used in the development of novel therapies for CAD.
The PennCath and MedStar studies were supported by the Cardiovascular Institute of the University of Pennsylvania, by the MedStar Health Research Institute at Washington Hospital Center and by a research grant from GlaxoSmithKline. The funding and support for the other cohorts contributing to the paper are described in the webappendix.
Definition of the genetic architecture of coronary artery disease (CAD) and myocardial infarction can provide substantial benefit through improved risk prediction and development of novel therapies. Recent genome-wide association studies (GWAS) provide promise, with identification of several novel loci for these disorders.1–6 However, only a small proportion of the inherited component has been identified.5 Atherosclerotic plaque rupture is the most common cause of myocardial infarction.7 Since all patients with plaque rupture or myocardial infarction have coronary atherosclerosis but only a few with coronary atherosclerosis develop myocardial infarction, unique factors—some genetic—are likely to predispose to plaque rupture or myocardial infarction in coronary atherosclerosis. Within clinically defined myocardial infarction, however, the mechanisms that drive events are unknown, such as those that cause progression of atherosclerosis, those that modulate plaque vulnerability, or factors that lead to arterial thrombosis.7 In fact, it is yet to be determined if identified loci for myocardial infarction5 contribute to initiation and progression of atherosclerosis or to plaque rupture and thrombosis in leading to myocardial infarction.
We report two GWAS of CAD designed to address the hypothesis that genetic factors predisposing to myocardial infarction in patients with coronary atherosclerosis are distinct from those that associate with the presence of coronary atherosclerosis. Unlike previous GWAS in this disease, we used coronary angiography in primary ascertainment of CAD phenotypes. This approach allows discrimination of risk alleles for plaque rupture and myocardial infarction from those for coronary atherosclerosis.
The webappendix shows a detailed description of design and clinical characteristics for study samples (pp 2–4). Our primary focus was on studies of angiographic CAD in patients of European ancestry. Discovery studies were PennCath and MedStar (webappendix p 2), which recruited patients before coronary angiography at the University of Pennsylvania Medical Center (Penn) and Washington Hospital Center (WHC), respectively. Selection of younger patients with angiographic CAD (mean age about 51±7 years) and older controls (mean age about 61±9 years) was aligned across these two studies, which used similar angiographic criteria and had similar numbers of cases of angiographic CAD with and without myocardial infarction.
PennCath is a hospital-based study of genes and biomarkers for CAD.1,5,8 Briefly, between July 1, 1998, and March 31, 2003, PennCath recruited 3815 consecutive patients undergoing cardiac catheterisation. All patients gave written informed consent. Clinical information was extracted from the medical records. Coronary angiograms were scored at the time of procedures. Blood was taken from patients in a fasting state, DNA and plasma were isolated, and lipoproteins and glucose were assayed. A nested case-control GWAS study of angiographic CAD was done in PennCath (n=1401 white patients). Controls (n=468) had no or minimum CAD (<10% stenosis of any vessel) on angiography and were aged older than 40 years for men and 45 years for women. Patients with angiographic CAD (n=933) had at least one coronary vessel with 50% or more stenosis, and men were aged 60 years or younger and women 65 years or younger. About half of patients with angiographic CAD presented with or had history of myocardial infarction (n=469).
MedStar is a hospital-based study designed for biomarker and genetic studies of CAD. Briefly, between Aug 1, 2004, and March 31, 2007, a case-control sample of patients with and without angiographic CAD undergoing cardiac catheterisation were recruited.5 All patients gave written informed consent and demographic, clinical, risk factor, and angiographic data were recorded. Study design of the case-control GWAS was similar to that for PennCath, with 1322 white patients composed of controls (n=447) and those with angiographic CAD (n=875). The patients with angiographic CAD were divided into those with myocardial infarction (n=421) and those without (n=454). Patients’ age at diagnosis of CAD was 55 years or younger for men and 60 years or younger for women. Controls were older than 45 years.
The webappendix (pp 2–4, 8, and 9) shows additional studies, including case-control subsets from the Wellcome Trust Case-Control Consortium (WTCCC) Study,3,4 the Ottawa Heart Genomics Study (OHGS),2,9 the Atherosclerotic Disease Vascular Function and Genetic Epidemiology study (ADVANCE),10 the Cadomics study,6,11 the Emory Genebank study,1 the Intermountain Heart Collaborative Study (IHCS),12 the Verona Heart Study (VHS),5,13 the Cleveland Clinic GeneBank Study,9 the Myocardial Infarction Genetics Consortium (MI-GEN),5 and the German myocardial infarction Family Studies.6
Detailed descriptions of genotyping, imputation, and quality control are shown in the webappendix (pp 4, 5, 10, and 11). Figure 1 shows the study design of the two multistage, case-control GWAS studies of phenotypes of angiographic CAD. A GWAS is a study in which genetic information is gathered for individuals with known phenotypes using a large array of genetic markers that represent common variation in the human genome. The aim is to map susceptibility genes through the detection of associations between these genetic markers and the phenotype of interest.14 First, we did a GWAS of patients with angiographic CAD compared with controls to identify loci for coronary atherosclerosis (study A, figure 1). We used a four-stage strategy similar to that used by Kathiresan and colleagues5 in their studies of myocardial infarction (study B, figure 1). Second, we did a GWAS study within patients with angiographic CAD comparing those who had myocardial infarction with those who did not to identify loci that predispose specifically to myocardial infarction. With a modification of a two-stage GWAS approach,15 we restricted the number of single nucleotide polymorphisms (SNPs) examined in stage 2, in view of modest power (webappendix p 5).
In stage 1 of study A, more than 2·4 million genotyped and imputed SNPs were studied. The union of top SNPs in PennCath, MedStar, and their meta-analysis (5425 SNPs with additive trend test p<1×10−3) were selected for stage-2 replication in analysis of existing genome-wide data for the WTCCC,3,4 OHGS,2,9 and ADVANCE10 studies (figure 1, webappendix pp 2, 3, and 8). After meta-analysis of stages 1 and 2, 25 novel SNPs (p <1×10−5) and ten SNPs from previous GWAS4,5 were selected for stage-3 wet-lab genotyping in the Emory GeneBank Study,1 Utah IHCS,12 and the VHS.5,13 The criteria for SNP selection at each stage were based on the approach of Kathiresan and colleagues.5 A significant finding was defined a priori by published criteria5,14 as p <5×10−8 in stages 1–3 combined as well as independent replication within stage-3 corrected for the number of SNP tests (n=35). For SNPs that reached genome-wide significance in stages 1–3, we undertook additional replication (figure 1, stage 4).9 Separately, we did an exploratory analysis of novel findings from study A in early-myocardial infarction studies consisting of MIGEN5 and GerMIFS6 (n=6256 myocardial infarction cases and 7711 controls).
For study B we assembled a six-study GWAS for an a-priori stage-1 meta-analysis of patients with angiographic CAD, comparing those with myocardial infarction with those without (figure 1, webappendix pp 2–4, and 9). Participants had evidence of angiographic CAD (by angiography in PennCath, MedStar, OHGS, CADomics and ADVANCE or by history of coronary revascularisation in WTCCC) and were classified for myocardial infarction as described.4–6 In most studies, patients with myocardial infarction were younger than those without and had broadly similar risk factors (webappendix p 9). In the stage-1 meta-analysis, between 1·2 and 2·4 million SNPs were examined. In stage 2, SNPs with p<5×10−6 in stage 1 (n=49) were interrogated in two Cleveland Clinic GeneBank studies9 with existing genome-wide data. A significant finding was defined a priori as p<5×10−8 in stages 1 and 2 combined as well as independent replication within stage 2 after Bonferroni correction for independent SNP tests (in stage 2, multiple SNPs were in strong linkage disequilibrium; therefore, we corrected for independent tests with the Nyholt method16). We also did a meta-analysis of all eight studies. Finally, we examined the association of published loci for myocardial infarction5 with angiographic CAD (in study A) as well as myocardial infarction in patients with angiographic CAD (in study B).
For genotyped SNPs, association was tested by logistic regression with the assumption of additive genetic effects with PLINK.17 For imputed SNPs, association was examined with SNPTEST, which can account for imputation uncertainty.18 We adjusted the effects of sex and age at diagnosis of CAD in all analyses. The estimated genomic control inflation factor lambda (λ),19 an indicator of potential population stratification, was calculated as the median of the test statistics divided by the median of χ2 distribution with one degree of freedom. For top SNPs identified in studies A and B, we also explored in PennCath the association with traditional risk factors such as diabetes, smoking, hypertension, hyperlipidaemia, and assessed top SNP associations with phenotypes of angiographic CAD after controlling for these risk factors. Finally, for top SNPs identified in study B, we used PennCath samples to determine the associations of these SNPs with myocardial infarction in patients with angiographic CAD after adjustment for Gensini score, a semi-quantitative estimate of the burden of angiographic CAD.20
Meta-analysis was done by a weighted Z-score method with METAL,21 which accounts for the direction of association relative to a consistent reference allele. In this method, p values for each study are first converted to a Z score. Then a weighted sum of Z scores is calculated, in which each statistic is weighted by the square root of the effective sample size for each study. The resulting sum is divided by the square root of the total effective sample size to obtain an overall Z statistic, which is used to assess the overall evidence for association. The reason for use of the effective sample size is to adjust for asymmetric case-control sample sizes. In accordance with de Bakker and colleagues,22 we used the non-centrality parameter for the given asymmetric case-control sample size, and then iteratively determined the effective (symmetric) case-control sample size that returns the same non-centrality parameter. To test for consistency of allelic effects across studies at the same SNP, we calculated two summary statistics of heterogeneity, Cochran’s Q statistic, which provides a test of heterogeneity of allelic effects at the test SNP, and an alternative I2 index, which quantifies heterogeneity in allelic effects across studies, over that expected by chance. The webappendix shows the power calculations (p 5).
In study B, SNPs at the ABO locus had genome-wide significant associations with myocardial infarction in patients with angiographic CAD. Therefore, in PennCath we inferred ABO blood groups and analysed blood group association with angiographic CAD phenotypes (webappendix pp 5 and 6). Stratified and conditional analyses were done in PennCath to examine whether the top ABO SNPs for myocardial infarction in patients with angiographic CAD were independent of ABO blood groups and distinct ABO SNPs.
Employees of GlaxoSmithKline contributed to study design, data interpretation, and editing of the report. The corresponding author had full access to the data and final responsibility for decision to submit for publication.
We examined loci for myocardial infarction, established through published GWAS,5 for their association with angiographic CAD in studies in which all cases and controls were defined by angiography (6886 patients with angiographic CAD and 3226 controls). The direction and strength of association with angiographic CAD for risk alleles (table 1) was largely consistent with published findings.4,5
Meta-analysis of stage-1–4 studies identified a novel genotyped SNP, rs1994016, on 15q25.1 that exceeded genome-wide significance for angiographic CAD (figure 2, table 2). This finding had modest heterogeneity across studies (Cochran’s Q=17·7, p=0·013; I2=0·60) with strong effects in initial studies, consistent with a so-called winners curse24 (figure 2, table 2). Within stages 1–3, rs1994016C reached genome-wide significance (table 2). In the same analysis, the published 9p21 rs4977574G allele (frequency about 0·56) had an OR of 1·35, p=2·98×10−27. Within the stage-3 wet-lab replications, rs1994016C had an OR of 1·14, p=4·97×10−4 (p=0·0174 after Bonferroni correction). In the stage-4 replication, the effect size was close to that seen in the other stages although the p value was 0·076, which probably indicates the small effective sample. In PennCath, rs1994016 was not associated with any risk factors for cardiovascular disease (data not shown) and its association with angiographic CAD was not attenuated after adjustment for risk factors (OR 1·32–1·35 before and after adjustment). The webappendix shows findings for all SNPs analysed in stages 1–3 (pp 12 and 13).
In separate analysis of MI-GEN and GerMIFS (early-onset myocardial infarction studies), the rs1994016C allele was not associated with myocardial infarction (OR 1·02, p=0·81; figure 2). This finding was unlikely to be due to low power because the analysed sample had more than 80% power to detect genotype relative-risk as low as 1·08.
The variant rs1994016 maps within intron 8 of ADAMTS7, a member of the family of disintegrin and metalloproteinase with thrombospondin motifs proteins.25 The webappendix shows the gene annotation and the linkage disequilibrium structure at this 15q25.1 region (p 23). The linkage disequilibrium is weak (r2<0·2) between rs1994016 and SNPs in any genes within 500 kb 3′ or 5′ of ADAMTS7, which suggests that ADAMTS7 is probably the atherosclerosis gene at this locus.
By contrast with angiographic CAD versus control, none of the published GWAS SNPs for myocardial infarction5 were significant (p<0·05) for myocardial infarction in patients with angiographic CAD (table 1); see for example, allele 9p21 rs4977574G). Our findings are unlikely to be attributable to low power, because this analysed sample had more than 80% power to detect genotype-relative risks from 1·07 to 1·20 for the reported effect sizes and allele frequencies at these loci.
In stage 1, 49 SNPs showed p<5×10−6 for myocardial infarction in patients with angiographic CAD. The top 11 SNPs mapped to one region on 9q34.2 within ABO, the blood group locus. Combined with stage-2 data, the association of multiple ABO SNPs, all in strong linkage disequilibrium (r2 >0·85), exceeded genome-wide significance (table 3; webappendix pp 14 and 24). For example, the OR of the G allele of the top genotyped SNP, rs612169, was 1·20 (p=3·66×10−8) for patients with angiographic CAD who had myocardial infarction whereas for the C allele at rs514659, the top imputed SNP, it was 1·21 (p=7·62×10−9, figure 3). Findings for rs514659 were consistent across studies in stages 1 and 2 (Cochran’s Q=4·8, p=0·69; I2=0, figure 3). Within stage-2 replication, the OR of the rs514659C allele was 1·24 (p=0·00145; p=0·0174 after Bonferroni correction for 12 independent tests). The webappendix shows results of the analysis of the combined stages 1 and 2 for all 49 SNPs selected in stage 1 (p 14).
We tested the association of risk alleles with specific angiographic CAD phenotypes in PennCath. Relative to controls, ABO rs514659C was related to angiographic CAD with myocardial infarction, but not to angiographic CAD without myocardial infarction (table 4), which suggests the ABO locus is related to myocardial infarction but not to coronary atherosclerosis. In PennCath, ABO SNPs were not associated with traditional risk factors (data not shown) and ABO associations with myocardial infarction in patients with angiographic CAD were not attenuated after risk factor adjustment (eg, rs514659, OR 1·38 before and after adjustment). Furthermore, in PennCath the signal for myocardial infarction in patients with angiographic CAD was not attenuated after adjusting for Gensini scores (eg, rs514659, OR 1·38 before and after adjustment). These findings suggest that results are not affected by slight differences in the burden of coronary atherosclerosis between angiographic CAD patients with with myocardial infarction and angiographic CAD patients without myocardial infarction.
Figure 4 and p 25 of webappendix show the gene annotation and linkage-disequilibrium structure of the GWAS region mapping to 9q34.2 as well as meta-analysis p values and recombination rates. The SNPs with strongest association lie in a linkage-disequilibrium block in intron 1 of the ABO locus. There is only modest linkage disequilibrium with SNPs (all r2<0·2) in other genes within 500 kb 5′ and 3′, suggesting that ABO is the myocardial infarction gene at this locus. The A allele of rs612169, our top genotyped SNP, tags the O blood group. Therefore, we inferred the main ABO blood groups using SNPs that tag ABO alleles26 (webappendix pp 5, 6, and 16) and examined their associations with CAD phenotypes in PennCath. Blood group A, B, and AB genotypes had greater odds of myocardial infarction in patients with angiographic CAD than did blood group O (table 4 and webappendix p 16). Subgroup analysis in PennCath revealed that, relative to controls, non-O blood groups had higher odds of angiographic CAD with myocardial infarction but not of angiographic CAD without myocardial infarction (table 4). This result suggests that ABO blood group O is protective against myocardial infarction in patients with angiographic CAD but is not related to coronary atherosclerosis. In a subgroup analysis of PennCath patients restricted to those with non-O blood groups, rs514659 was not associated with myocardial infarction in patients with angiographic CAD (OR 0·90 [0·59–1·38], p=0·63), suggesting that rs514659 has no effect independent of the ABO blood group allele.
We identified two loci for distinct CAD phenotypes, ADAMTS7, a novel locus for angiographic CAD but not myocardial infarction, and ABO, a gene for myocardial infarction in patients with angiographic CAD, but not for angiographic CAD itself. Further, our data suggest that the ABO GWAS signal for myocardial infarction in patients with angiographic CAD is mediated by the glycotransferase-deficient isoform that encodes the ABO blood group O phenotype.
Clinical CAD phenotypes are heritable but highly complex. The association of several published loci for myocardial infarction5 might be mediated by diverse pathological processes including those that promote atherosclerosis, precipitate plaque rupture, or facilitate arterial thrombosis. Our use of coronary angiography reduced heterogeneity in coronary atherosclerosis within patients with CAD while allowing discrimination of risk alleles for plaque rupture or myocardial infarction from those for atherosclerosis. Although most published loci for myocardial infarction had significant signals for angiographic CAD compared with controls, none were associated with myocardial infarction in patients with angiographic CAD. This finding suggests that these loci relate to myocardial infarction indirectly via coronary atherosclerosis rather than having a specific role in vulnerable plaque and myocardial infarction. In fact, independent studies support this concept for the 9p21 locus. Consistent with our data, Horne and colleagues12 showed that this locus did not predict incident or prevalent myocardial infarction in patients with CAD but was strongly associated with the presence of angiographic CAD versus controls.
Our discovery of ADAMTS7 as a novel locus for CAD might have been facilitated by use of coronary angiography because, unlike clinically defined cases, the definition of angiographic CAD required a pre-specified burden of coronary atherosclerosis. All ADAMTS genes have a similar domain structure, consisting of a preproregion, a reprolysin-type catalytic domain, a disintegrin-like domain, a thrombospondin type-1 module, a cysteine-rich domain, a spacer domain, and a COOH-terminal thrombospondin type-1 module. ADAMTS7 degrades cartilage oligomeric matrix protein and has been implicated in inflammatory arthritis and bone growth. Overexpression of ADAMTS7 accelerates migration of vascular smooth muscle cells in vitro and exacerbates neointimal thickening after carotid artery injury in vivo, perhaps through degradation of cartilage oligomeric matrix protein.27 These data implicate ADAMTS7 in the proliferative response to vascular injury, a process that has parallels to the progressive phase of atherosclerosis.7 These mechanistic findings coincide with the lack of association of ADAMTS7 SNPs with early-onset myocardial infarction. Together they raise the provocative possibility that some proteins, such as ADAMTS7, could increase plaque size but not affect plaque stability. Overall, ADAMTS7 might be a novel therapeutic target for progression of atherosclerosis but seems less likely to be one for prevention of myocardial infarction in high-risk patients.
Discovery of ABO as the top locus for myocardial infarction in patients with angiographic CAD is notable, in view of decades of work suggesting a relation between ABO blood groups and both thrombosis and coronary heart disease.28,29 The ABO gene encodes proteins (transferase A, a 1-3-N-acetylgalactosaminyltransferase; transferase B, a 1-3-galactosyltransferase) related to the ABO blood group system.30 Blood group O is caused by a deletion of guanine-258 near the N-terminus of the protein. This deletion causes a frameshift, which results in translation of a protein with no glycosyltransferase activity.30 In a meta-analysis,29 Wu and colleagues reported ORs for non-O relative to the O blood group of 1·79 (1·56–2·05) for venous thrombo embolism, 1·25 (1·14–1·36) for myocardial infarction, but only 1·03 (0·89–1·19) for angina.29 Ketch and colleagues31 reported that patients with non-O blood groups had higher thrombus burden despite less extensive coronary atherosclerosis at the time of acute myocardial infarction. These data, coupled with our genetic findings, strongly suggest a primary relation of non-O ABO glycotransferase activity with coronary thrombosis rather than atherosclerosis. ABO-related thrombosis is thought to be mediated by ABO carbohydrate-modification of von Willebrand Factor (VWF) resulting in impaired proteolysis and higher circulating von Willebrand Factor and Factor VIII.32
Tregouet and co-workers identified ABO as the most significant locus in a GWAS of venous thrombo-embolism.33 The top ABO SNPs for venous thromboembolism, rs657152 and rs505922, have strong associations with myocardial infarction in patients with angiographic CAD in our data (eg, rs505922 p=1·032×10−8) and are in strong linkage disequilibrium with our top ABO SNP signals for myocardial infarction (eg, r2 1·0 with rs514659 (figure 4, table 3, webappendix p 6). We note that the top SNPs (rs687621 and rs687289) in a GWAS34 of plasma von Willebrand Factor and Factor VIII are in complete linkage disequilibrium with rs514659 and blood group O and also reach genome-wide significance for myocardial infarction in patients with angiographic CAD (table 3). Thus, common ABO genetic variation, linked to blood group O, reduced glycotransferase activity and lower circulating von Willebrand Factor and Factor VIII, lowers risk of myocardial infarction in the setting of angiographic CAD, while also protecting against venous thromboembolism.
The relation between ABO and atherosclerotic cardiovascular disease, however, might be more complex than modulation of thrombosis. Other GWAS also identified ABO as a locus for low-density lipoprotein (LDL-C),35 type-2 diabetes,26 inflammatory risk biomarkers E-selectin, P-selectin, and sol-ICAM126,36–38 as well as angiotensin-converting enzyme39 (figure 4). Indeed, ABO blood group associations with plasma cholesterol were described several decades ago. In PennCath, however, ABO SNP associations with myocardial infarction in patients with angiographic CAD were not attenuated by adjustment for ABO SNPs related to LDL-C, ICAM-1, and E-selectin (webappendix p 6). Taken together, these factors suggest that ABO might modulate various distinct pathways related to cardiovascular risk factors, atherosclerosis, and thrombosis.
Our study has potential limitations. The multistudy design might have introduced selection bias and confounding. The absence of genomic control inflation, however, argues against confounding caused by genetic differences in source populations. Angiography cannot detect early subclinical atherosclerosis in controls resulting in misclassification of controls as free of CAD. Misclassification is potentially a greater drawback in our study within patients with angiographic CAD, in which those without myocardial infarction might subsequently develop myocardial infarction. Patients with angiographic CAD who had myocardial infarction, however, tended to be younger than those who did not have myocardial infarction, despite having broadly similar risk factors. This finding suggests that additional factors beyond age and traditional risk must contribute to myocardial infarction among patients with angiographic CAD. Overall, heterogeneity and misclassification would tend to bias to wards the null, would not affect our novel findings, but would limit the power for additional discoveries.
Our results indicate that specific genetic variants predispose to the development of coronary atherosclerosis whereas others predispose to subsequent plaque rupture and acute myocardial infarction. Further, many published loci for myocardial infarction are likely to relate to the initiation and progression of coronary atherosclerosis rather than having a specific role in vulnerable plaque and myocardial infarction. Translation of GWAS discoveries for CAD into prognostic and therapeutic benefit will need greater insights into the relation between each locus and the phenotypes of atherosclerosis, plaque rupture, and thrombosis.
All acknowledgements are provided online in pp 19–22 of the webappendix.
ContributionsMPR, ML, SEE, SK, NJS, and DJR contributed to study design. JFF, IMS, NNM, MSB, JMD, BDH, AFRS, TLA, PSW, RSP, PLN, NM, DG, AAQ, JLA, JE, ASH, HS, TQ, SB, SLH, RR, DJR, and NJS contributed to data collection and performed research. ML, JH, MPR, JFF, JRT, HA, NM, CWK, BDH, AFRS, TLA, JE, TQ, SB, SLH, RR, SK, DJR, and NJS contributed to data analysis and interpreted results. MPR, ML, JFF, and DJR wrote drafts of the manuscript. MPR, ML, JFF, MSB, JMD, CWK, BDH, AFRS, TLA, PSW, HA, PLN, RSP, DG, AAQ, JLA, JE, ASH, HS, TQ, SB, SLH, RR, SK, NJS, SEE, and DJR revised and reviewed the final manuscript.
Conflicts of interest
MPR, MSB, JMD, SEE, and DJR received research grant support from GlaxoSmithKline. CWK was an employee of GlaxoSmithKline at the time of the study. ML, JH, JFF, IMS, NNM, JRT, BDH, AFRS, TLA, PSW, HA, PLN, RSP, NM, DG, AAQ, JLA, JE, ASH, HS, TQ, SB, SLH, RR, SK, and NJS declare that they have no conflicts of interest.