Polygenic risk scores identify heterogeneity in asthma and chronic obstructive pulmonary disease

Abstract

Rationale:

Asthma and chronic obstructive pulmonary disease (COPD) have distinct and overlapping genetic and clinical features.

Objectives:

We hypothesized that polygenic risk scores (PRSs) for asthma (PRS_Asthma) and spirometry (FEV₁ and FEV₁/FVC; PRS_spiro) would demonstrate differential associations with asthma, COPD, and asthma-COPD overlap (ACO).

Methods:

We developed and tested two asthma PRSs and applied the higher performing PRS_Asthma and a previously-published PRS_spiro to research (COPDGene and CAMP, with spirometry) and electronic-health record (EHR)-based (MGB Biobank and GERA) studies. We assessed the association of PRSs with COPD and asthma using modified random and binary effects meta-analyses, and ACO and asthma exacerbations in specific cohorts. Models were adjusted for confounders and genetic ancestry.

Measurements and Main Results:

In meta-analyses of 102,477 participants, the PRS_Asthma (OR per SD 1.16 [95% CI: 1.14–1.19]) and PRS_spiro (OR per SD 1.19 [95% CI: 1.17–1.22]) both predicted asthma, while the PRS_spiro predicted COPD (OR per SD 1.25 [95% CI: 1.21–1.30]). However, results differed by cohort. The PRS_spiro was not associated with COPD in GERA and MGB. In COPDGene, the PRS_Asthma (OR per SD: Whites: 1.3; African Americans (AA): 1.2) and PRS_spiro (OR per SD: Whites: 2.2; AA: 1.6) were both associated with ACO. In GERA, the PRS_Asthma was associated with asthma exacerbations (OR 1.18) in whites; the PRS_spiro was associated with asthma exacerbations in white, LatinX, and East Asian participants.

Conclusions:

Polygenic risk scores for asthma and spirometry are both associated with asthma-COPD overlap and asthma exacerbations. Genetic prediction performance differs in research versus EHR-based cohorts.

Capsule Summary:

Genetic variants can predict COPD, asthma, asthma-COPD overlap, and asthma exacerbations, but accuracy varies. Accuracy is high for spirometry-defined COPD, but lower for asthma and asthma exacerbations, and low for health record-defined COPD.

Introduction

Asthma and chronic obstructive pulmonary disease (COPD) are characterized by airflow limitation and are associated with significant global morbidity and mortality (1–3). While the ‘classic’ presentation of these diseases - for example, a child with atopy and asthma, or an older adult smoker with no history of asthma - may be easy to classify, in practice, there is a substantial amount of overlap. Some of this overlap is likely due to patients with shared features of both diseases, also known as asthma-COPD overlap (ACO), which has been associated with worse outcomes than either condition alone (4–7). However, some overlap may also be due to challenges in classifying disease. Asthma and COPD identified by medical records are often clinical diagnoses without spirometry confirmation, and both diseases suffer from high rates of under- and over-diagnosis (8–10).

The risks for developing either disease are influenced by genetic and environmental factors (4–7). The shared and divergent features between asthma and COPD might be partially explained by genetics. The ‘Dutch hypothesis’ suggests that asthma and COPD share common genetic origins and that certain individuals with asthma early in life may go on to develop COPD (11,12). Genetic association studies have identified numerous loci associated with asthma (13–15), spirometry measures (16,17), COPD (18), and ACO (4,19). While individual genome wide association study (GWAS) variants tend to have small effect sizes and explain a small amount of phenotypic variability, aggregating thousands to millions of variants into a polygenic risk score (PRS) can explain more phenotypic variability and improve complex trait prediction (20). PRSs have been developed for asthma (21,22) and spirometry measures (23), and for the latter, have been used to understand COPD disease heterogeneity and susceptibility (24–26).

We hypothesized that PRSs for asthma and spirometry would be associated with asthma and COPD diagnoses in both research and ‘real-life’ electronic health record (EHR)-based clinical cohorts, but that these associations would differ based on the cohort due to differences in ascertainment. We further hypothesized that these scores could be used to dissect how the genetic risks of these traits contribute to phenotypic heterogeneity with respect to ACO, markers of atopy, individuals who were diagnosed with asthma early in life who later developed COPD with emphysema, asthma exacerbations, and medication use. In a prior study, we used summary statistics from a GWAS of over 500,000 individuals (16) to construct PRSs for FEV₁ and FEV₁/FVC(23). We summed these two PRSs to create a PRS for COPD, which was predictive of COPD (odds ratio per standard deviation, 1.8) in over 50,000 external participants from 9 cohorts. The PRS was also associated with emphysema, airway wall thickness, and reduced lung function growth (23). In the present study, we refer to this PRS as the PRS_spiro to emphasize its derivation when assessing its relation to asthma and asthma-COPD overlap.

Methods

Study populations

All study participants provided informed consent and protocols were approved by local Ethics Committees and Institutional Review Boards. Additional information regarding cohorts, sample collection, and genotyping details can be found in the Supplement.

CAMP

We included individuals from the Childhood Asthma Management Program (CAMP) with genetic and spirometry data 23 years after study enrollment (27–29). Briefly, CAMP was a randomized trial of anti-inflammatory treatments in non-Hispanic white (NHW) children with asthma ages 5 to 12 years at enrollment who were followed up for 13 years and had low attrition (≤20%). Children followed into adulthood were diagnosed with COPD by applying a lower limit of normal criterion to post-bronchodilator spirometry measures (29,30).

COPDGene

We included participants from the Genetic Epidemiology of COPD (COPDGene) study (31). In this study, 10,198 NHW and African American (AA) participants aged 45 to 80 years with ≥ 10 pack-years of smoking were recruited from 21 clinical centers across the United States. Anthropometric and survey data, spirometry, CT imaging, and blood were collected at baseline and at the 5- and 10-year follow up visits. From this research study cohort, we included participants with asthma, COPD, or ACO phenotype data.

MGB Biobank

We included participants from Mass General Brigham (MGB) Biobank, which contains over 117,000 individuals with corresponding survey and electronic health record (EHR) data since 2009 (32). From this clinical cohort, we included individuals of European ancestry with genotype and phenotype data for asthma and COPD.

GERA

We included participants from the Kaiser Permanente Northern California (KPNC) Genetic Epidemiology Research on Adult Health and Aging (GERA) cohort (33–35). This clinical cohort includes over 100,000 individuals with EHR and genotyping data. We included participants who were at least 21 years of age, and who had self-reported race/ethnicity, asthma or COPD phenotype information, and genotyping data.

Polygenic risk scores: asthma

For the PRS_Asthma, we derived scores from two genome-wide association studies (GWASs), compared their performances in testing cohorts, and used the better performing score in our analysis. The first GWAS was from Han et al. (15), and included 88,486 cases and 447,859 controls from the UK Biobank and Trans-National Asthma Genetic Consortium (TAGC) (hereafter, referred to as the UKB/TAGC GWAS). TAGC was performed before UK Biobank and did not include overlapping participants. The second GWAS was from the Global Biobank Meta-analysis Initiative (GBMI), a collaborative meta-analysis of 19 biobanks across 4 continents representing 2.1 million individuals (14); for the asthma GWAS, 7 ancestry groups comprising 153,763 asthma cases and 1,647,022 controls were used. The GBMI meta-analysis included MGB Biobank participants. We used logistic liability R² calculations (36) to compare scores derived from these two studies because one GWAS utilized fewer participants but included more well-phenotyped asthma cases (UKB/TAGC), while the larger study (GBMI) was based primarily on administrative billing/coding practices.

We then applied lassosum (37), a penalized regression-based method that accounts for linkage disequilibrium (LD), calculating LD using European ancestry individuals from the UK Biobank (38). As was previously performed in the development of the PRS_spiro (23), we used the Genetics of COPD (GenKOLS) case-control study from Norway (39) to tune hyperparameters for the PRS_Asthma.

Polygenic risk scores: spirometry

The PRS_spiro (a.k.a. COPD PRS) was calculated as previously described (23), and is also detailed in the Supplement. All PRSs were scaled and centered and results are reported per standard deviation (SD) of the risk score.

Statistical analyses

Overview of study design

After identifying the highest-performing PRS_Asthma based on the highest average R² for asthma across cohorts, we calculated the PRS_Asthma and PRS_spiro in each cohort. We meta-analyzed the association of each PRS with asthma and COPD. We also used the PRSs to dissect how genetic risk for asthma and spirometry account for heterogeneity in asthma, COPD, and ACO by applying regression frameworks to examine the associations of PRSs with particular phenotypes available in cohorts. Within each cohort, we performed analyses in each racial/ethnic group, adjusting for ancestry within each group.

Outcomes

The primary outcomes were diagnosis of asthma or COPD. In CAMP, all participants were diagnosed with asthma as children, and COPD was defined as those with FEV₁/FVC below the lower limit of normal based on smoothed RLOWESS spirometry curves at 23 years of follow up (29,30). For the primary analysis of COPD and asthma, in all cohorts except CAMP, asthma participants were excluded from COPD phenotype definitions and COPD participants were excluded from asthma phenotype definitions. In COPDGene, COPD was defined according to Global Initiative for Obstructive Lung Disease (GOLD) (40) criteria for moderate-to-severe obstruction (GOLD 2–4: FEV₁ % predicted < 80% and FEV₁/FVC < 0.7) without a history of asthma. Asthma was defined as having a self-reported history of asthma with FEV₁/FVC ≥ 0.7. As in prior studies (4,5), ACO was defined as GOLD 2–4 participants with a self-reported history of asthma that was diagnosed by a physician before age 40. In MGB Biobank, asthma and COPD were defined using machine learning algorithms (see Supplement for details). Briefly, MGB COPD and asthma phenotypes were identified by applying least absolute shrinkage and selection operator (LASSO) (41) penalized regression to EHR information including demographics, smoking history, diagnosis codes, and medications, but without spirometry. In GERA, the asthma case definition was previously used in a multi-ancestry asthma GWAS (35) and included individuals with self-report of asthma or a doctor diagnosis of asthma or an asthma exacerbation and no COPD diagnosis. COPD diagnosis was based on the presence of an International Classification of Diseases (ICD)-9 codes and the absence of diagnosis codes for asthma (35).

We examined multiple secondary outcomes based on clinician input for phenotypes relevant to asthma and COPD heterogeneity, which are detailed in the Supplement.

Model specifications and performance evaluations

We considered the PRS_Asthma and PRS_spiro as potential predictors of the above outcomes in multivariable linear and logistic regressions, as appropriate. All models were adjusted for age, sex, current smoking status (except in CAMP), pack-years of smoking (available in COPDGene and MGB Biobank), and principal components of genetic ancestry. For CT imaging outcomes, we additionally adjusted models for scanner type. For ‘asthma to COPD’ (COPDGene only, defined in Supplement), we performed an additional analysis adjusting only for sex to model the available clinical variables in a child with asthma. For primary outcomes, we performed binary effects meta-analyses using metasoft, and also performed inverse-variance fixed and random effects meta-analyses and created forest plots using the meta R package (42). Due to the anticipated study heterogeneity, we used binary effects meta-analysis as our primary result for hypothesis testing (p-values).

All analyses were done in R version 4.0.3 (www.r-project.org). Normality for continuous variables was assessed by visual inspection of histograms. Results are shown as mean ± standard deviation or median [interquartile range], as appropriate. Differences in continuous variables were assessed with Student t-tests or Wilcoxon tests. Categorical variables were compared by ANOVA or Kruskal-Wallis tests, as appropriate. P-values below a Bonferroni-corrected 𝖺 were considered significant and those below 0.05 were considered nominally significant.

Results

Asthma and Lung Function Polygenic Risk Scores

A schematic of the study design is shown in Figure 1. We included 102,477 participants with genetic and phenotype data for asthma and COPD from four cohorts (Table 1). To develop a PRS_Asthma, we input UKB/TAGC (15) and GBMI (14) asthma GWAS summary statistics into a penalized regression framework (37) to develop two asthma PRSs. We compared the performance of these scores in external testing cohorts (Table E1) and found that the UKB/TAGC PRS_Asthma (R² 0.021) performed better than the GBMI PRS_Asthma (R² 0.011) on the logistic liability scale; hereafter, PRS_Asthma will refer to the UKB/TAGC score. The PRS_Asthma contained 563,117 variants with minimal variant dropout rates (Table E2) and the selected shrinkage was 0.2 and 𝝀 was 0.00546. The PRS_spiro was based on GWASs of lung function (R² ~0.30), and we calculated the previously published score (23) in testing cohorts. To examine the overlap between PRS_Asthma and PRS_spiro, we calculated the number of shared variants (Table E3) and the correlation of the PRSs across cohorts (Table E4). In LD score regression, the genetic correlation between asthma and FEV₁ was −0.295, and between asthma and FEV₁/FVC was - 0.319.

Figure 1:

Schematic of study design. COPDGene = Genetic Epidemiology of COPD study. MGB = Mass General Brigham. GERA = Genetic Epidemiology Research on Aging. CAMP = Childhood Asthma Management Program. COPD = chronic obstructive pulmonary disease. GenKOLS = Genetics of COPD study in Norway. GBMI = Global Biobank Meta-analysis Initiative. GWAS = Genome-wide association study. ACO = asthma-COPD overlap. PRS = polygenic risk score. NHW = non-Hispanic white. AA=African American.

Table 1.

Characteristics of study participants.

Characteristics	Overall	COPDGene NHW	COPDGene AA	MGB Biobank	CAMP	GERA East Asian	GERA AA	GERA white European	GERA LatinX
n	102477	5310	2466	10898	574	6653	2864	67819	5893
Age in years (mean (SD))	60.43 (13.63)	62.22 (8.81)	54.91 (7.43)	56.70 (15.99)	25.98 (1.80)	55.56 (14.11)	58.33 (13.40)	62.44 (12.79)	54.85 (14.25)
Sex (No. % female)	57807 (56.4)	2526 (47.6)	1045 (42.4)	6092 (55.9)	351 (61.1)	3765 (56.6)	1635 (57.1)	38798 (57.2)	3595 (61.0)
Race (No. %)
non-Hispanic white or European	84597 (82.6)	5310 (100.0)	0 (0.0)	10898 (100.0)	574 (100.0)	0 (0.0)	0 (0.0)	67819 (100.0)	0 (0.0)
African American	5331 (5.2)	0 (0.0)	2466 (100.0)	0 (0.0)	0 (0.0)	0 (0.0)	2864 (100.0)	0 (0.0)	0 (0.0)
Hispanic	5893 (5.8)	0 (0.0)	0 (0.0)	0 (0.0)	0 (0.0)	0 (0.0)	0 (0.0)	0 (0.0)	5893 (100.0)
Other	6654 (6.5)	0 (0.0)	0 (0.0)	0 (0.0)	0 (0.0)	6653 (100.0)	0 (0.0)	0 (0.0)	0 (0.0)
Pack-years of smoking (median [IQR])	14.10 [0.00, 38.70]	42.20 [30.60, 60.00]	34.75 [23.42, 46.80]	0.00 [0.00, 7.50]	NA	NA	NA	NA	NA
Current smoker (No. %)	8533 (8.7)	1994 (37.6)	1937 (78.5)	595 (5.5)	0 (NA)	242 (3.8)	241 (8.9)	3178 (4.9)	346 (6.2)
Asthma (No. %)	20957 (21.5)	208 (10.3)	230 (15.3)	234 (2.1)	574 (100)	1513 (22.7)	753 (26.3)	16440 (24.2)	1579 (26.8)
Asthma-COPD overlap (No. %)	854 (0.83)	581 (0.57)	273 (0.27)	0 (NA)	0 (NA)	0 (NA)	0 (NA)	0 (NA)	0 (NA)
COPD (No. %)	7639 (7.5)	3065 (59.2)	910 (36.9)	53 (0.5)	133 (23.2)	223 (3.4)	131 (4.6)	2885 (4.3)	239 (4.1)

COPDGene = Genetic Epidemiology of COPD study. MGB = Mass General Brigham. GERA = Genetic Epidemiology Research on Aging. CAMP = Childhood Asthma Management Program. COPD = chronic obstructive pulmonary disease. ACO = asthma-COPD overlap. NHW = non-Hispanic white. AA=African American. Asthma, COPD, and asthma-COPD overlap participants were mutually exclusive.

Association of Polygenic Scores with Asthma and COPD

The associations of the PRS_Asthma and PRS_spiro with asthma and COPD phenotypes across cohorts is shown in Figure 2 and Table 2. The PRS_Asthma was significantly associated with asthma (OR per SD 1.16 [95% CI: 1.14 to 1.19]), but not COPD (OR 0.99 [95% CI: 0.96 to 1.03]). The PRS_spiro was significantly associated with both asthma (1.19 [95% CI: 1.17 to 1.22]) and COPD (OR 1.25 [95% CI: 1.21 to 1.30]). However, while the effect of the PRS_Asthma on asthma diagnosis was seen in all cohorts (M-value, or posterior probability of association > 0.8), the other results had substantial heterogeneity. The effect of the PRS_spiro on asthma was ambiguous in COPDGene AA and MGB Biobank (M-value ~0.4), while the association of the PRS_spiro with COPD was only observed in research cohorts that defined COPD based on spirometry. Stratified analyses of males and females and former and current smokers yielded similar results (Figures E1–E4, Table E5), except in CAMP, in which we observed that the PRS_spiro was primarily associated with COPD in males (Figures E1 and E2).

Figure 2:

Inverse-variance random effects meta-analyses of multivariable logistic regressions of asthma (PRS_Asthma) and spirometry (PRS_spiro) PRSs with asthma and COPD outcomes were performed. Models were adjusted for age, sex, pack-years of smoking (when available), current smoking status, and principal components of genetic ancestry. See the legend of Figure 1 for abbreviations.

Table 2.

Modified random effects (random effects 2) and binary effects meta-analyses of multivariable logistic regression models testing the association of asthma (PRS_Asthma) and spirometry (PRS_spiro) PRSs with asthma and COPD diagnosis.

outcome	PRS	p (random effects 2)	p (binary effects)	$I^{\wedge }2$	Q	p (Q)	M-values
							COPDGene NHW	COPDGene AA	MGB Biobank	GERA white European	GERA African American	GERA East Asian	GERA LatinX	CAMP
Asthma (n=93,738)	PRSAsthma PRSspiro	2.07E-46 3.68E-70	7.80E-46 4.82E-71	42.4 54.7	10.4 13.2	0.108 0.0394	1 0.959	0.996 0.411	0.808 0.436	1 1	0.993 0.991	0.995 0.946	0.999 1	NA NA
COPD (n=98,436)	PRSAsthma PRSspiro	0.331 1.11E-104	0.58 3.78E-101	15.2 97.9	8.25 335	0.311 2.48E-68	0.224 1	0.242 1	0.326 0.001	0.34 0	0.305 0	0.303 0	0.328 0	0.318 0.043

Models were adjusted for age, sex, pack-years of smoking (when available), current smoking status, and principal components of genetic ancestry. These meta-analysis approaches enhance statistical power when there is marked study heterogeneity. In modified random effects analyses, the null hypothesis assumes no study heterogeneity. In binary effects analyses, each study is given a value to predict whether the study has a significant association (M-values), and the weighted Z scores are summed and a p-value is calculated. Q = Cochrane’s Q statistic. n=number of participants included in analysis.

Polygenic risk scores lend insight into asthma and COPD heterogeneity

In COPDGene, the PRS_Asthma (OR per SD: NHW: 1.3 [95% CI: 1.2–1.5]; AA: 1.2 [95% CI: 1.0–1.4]) and PRS_spiro (OR per SD: NHW: 2.2 [95% CI: 2.0 to 2.5]; AA: 1.6 [95% CI: 1.3–1.8]) were both significantly associated with ACO (Figure 3). The PRS_Asthma was nominally associated with CT airway thickness measures (WA% and Pi10) in COPDGene AA participants (Figure E5), but not in NHWs, nor was the PRS_Asthma associated with CT emphysema measures. Characteristics of ‘asthma to COPD’ participants (see Methods for further details) versus those with asthma diagnosed before age 40 years is shown in Table E6. In COPDGene NHW participants, the PRS_spiro was associated with ‘asthma to COPD’ (OR 1.59 [95% CI: 1.15 to 2.2], p = 0.0053) (Table 3). Compared to a clinical model including age, sex, and smoking variables, adding the PRSs increased the AUC from 0.72 to 0.77 (p=0.027) with a net reclassification index of 2.1% (95% CI: 0.68% to 3.8%). We did not observe these results in COPDGene AA participants (PRS_Asthma OR 1.27 [95% CI: 0.77 to 2.11], p = 0.4; PRS_spiro OR 0.932 [95% CI: 0.58 to 1.5], p = 0.8). We observed similar results when adjusting only for sex, the clinical variable that would be available in a child with asthma (Table E7).

Figure 3.

The association of asthma and spirometry PRSs with asthma phenotypes and ACO are shown. Asthma < 40 = doctor diagnosis of asthma before age 40 years. Asthma < 16 = doctor diagnosis of asthma before age 16 years. ACO = asthma-COPD overlap. COPDGene = Genetic Epidemiology of COPD study. PRS = polygenic risk score. NHW = non-Hispanic white. AA=African American. * indicates p-values below a Bonferroni-corrected significance level of 0.003125.

Table 3.

Multivariable associations of asthma (PRS_Asthma) and lung function (PRS_spiro) PRSs with ‘asthma to COPD’ phenotype (i.e. being diagnosed with asthma before age 40 and having moderate-to-severe obstruction (GOLD 2–4) with greater than 5% emphysema (% LAA < −950 HU) at the time of study entry into COPDGene). Models were adjusted for age, sex, pack-years of smoking, current smoking status, and principal components of genetic ancestry.

‘Asthma to COPD’ phenotype	PRSAsthma		PRSspiro
cohort	OR (95% CI)	p	OR (95% CI)	p
COPDGene NHW (n=352)	1.05 (0.775 – 1.43)	0.74	1.59 (1.15 – 2.2)	0.0053*
COPDGene AA (n=276)	1.27 (0.77 – 2.11)	0.35	0.932 (0.577 – 1.5)	0.77

In COPDGene NHW participants, a model with asthma and lung function PRSs had an AUC of 0.67, a clinical model (age, sex, height, pack-years of smoking, current smoking status) had an AUC of 0.72, and a combined model with PRSs and clinical factors had an AUC of 0.77 (p [AUC combined vs. AUC clinical model] = 0.027). In COPDGene AA participants, a model with asthma and lung function PRSs had an AUC of 0.68, a clinical model had an AUC of 0.78, and a combined model with PRSs and clinical factors had an AUC of 0.79 (p [AUC combined vs. AUC clinical model] = 0.9). GOLD = Global Initiative for Obstructive Lung Disease. LAA = low attenuation area. HU = Hounsfeld units. AUC = area-under-the-receiver-operating-characteristic-curve. n=number of participants included in analysis.

^*indicates p-value below Bonferroni threshold of 0.0083. n=number of participants included in analysis.

We additionally investigated whether the PRS_Asthma was associated with eosinophil counts or IgE levels in COPD participants (not excluding asthma). The PRS_Asthma was nominally associated with a 5% higher eosinophil count and higher IgE levels in male NHW COPD participants (Tables E8 and E9). In CAMP, the PRS_Asthma was not associated with a reduced lung function growth pattern (OR 0.95 [95% CI: 0.77 to 1.2], p=0.64).

Polygenic risk scores are associated with asthma exacerbations

We leveraged the large EHR data associated with GERA to examine the association of PRSs with asthma exacerbations and medication class use. The PRS_spiro was associated with an increased odds of asthma exacerbations in GERA East Asian (OR per SD 1.13 [95% CI: 1.03 to 1.23]), GERA white European (OR per SD 1.18 [95% CI: 1.15 to 1.22]) and GERA LatinX (OR per SD 1.13 [95% CI: 1.04 to 1.24]) participants (Table 4 and Table E10). The PRS_Asthma was also associated with asthma exacerbations in GERA white European (OR per SD: 1.1 [95% CI: 1.07 to 1.14]) participants. Model AUCs ranged from 0.57–0.61 across GERA cohorts.

Table 4.

Multivariable associations of asthma (PRS_Asthma) and spirometry (PRS_spiro) PRSs with asthma exacerbations in GERA.

Exacerbations		PRSAsthma		PRSspiro
cohort	n	OR (95% CI)	p	OR (95% CI)	p	AUC
GERA East Asian	6,164	0.999 (0.914 – 1.09)	0.99	1.13 (1.03 – 1.23)	0.01*	0.61
GERA African American	2,624	1.09 (0.972 – 1.21)	0.15	1.13 (0.992 – 1.28)	0.067	0.57
GERA white European	63,392	1.1 (1.07 – 1.14)	4.4e-11*	1.18 (1.15 – 1.22)	5.2e-35*	0.58
GERA LatinX	5,408	1.05 (0.969 – 1.14)	0.23	1.13 (1.04 – 1.24)	0.005*	0.57

Models were adjusted for age, sex, current smoking status, and principal components of genetic ancestry.

^*indicates p-values below a Bonferroni-corrected threshold of 0.0125. AUC = area-under-the-receiver-operating-characteristic-curve. n=number of participants included in analysis.

Discussion

In this study of over 100,000 participants from two research and two EHR-based clinical cohorts, we used polygenic risk scores (PRSs) for asthma and spirometry to identify associations with asthma, COPD, ACO, and related phenotypes. Asthma diagnosis was predicted by both PRSs, while COPD diagnosis was predicted by a PRS for spirometry, but only in cohorts where spirometry was used to define COPD. ACO individuals exhibited a higher predicted genetic risk for both asthma and low spirometry measures, and a higher genetic risk for low lung function predicted which individuals diagnosed with asthma before age 40 were likely to develop moderate-to-severe COPD with emphysema. We observed that both PRSs were predictive of asthma exacerbations in European ancestry individuals. These results suggest that the genetic risks for asthma and low lung function are important for ACO and asthma exacerbation risks, as well as raise questions about disease diagnoses and shared biology.

Previous studies have identified genetic correlations between asthma and COPD to be ~0.38 (19,43) to as high as 0.67 (44). While we observed genetic correlations between asthma with FEV₁ and asthma with FEV₁/FVC (~ −0.30), the PRS_Asthma and PRS_spiro were only weakly correlated. These results may be due to variations in phenotypes and cohorts used to calculate this genetic correlation, and the inability of PRSs to completely capture heritability. To calculate our PRS_Asthma we used lassosum and observed improved performance in GERA African Americans compared to our prior work (22). The PRS_Asthma was associated with asthma across ancestry groups but was not associated with COPD. While some of this result may be due to the relatively low power of the PRS_Asthma, it also suggests that, while the genetic determinants of asthma may overlap with those of spirometry, the PRS_Asthma may reflect genetic risk for asthma that is not largely driven by the genetic risk for low spirometry measures.

Our findings highlight the challenges in translating genetic prediction tools derived from research cohorts to EHR-based clinical cohorts. The PRS_spiro was derived from GWASs of FEV₁ and FEV₁/FVC and has been validated in 9 well-phenotyped case-control and population-based cohorts (23). In the current study, the PRS_spiro was tested for association with COPD subjects without a history of asthma and also tested in real-world EHR-based cohorts. The PRS_spiro was associated with COPD in research studies (COPDGene and CAMP where COPD was defined by spirometry), but not the EHR-based cohorts where COPD was defined by billing codes. These findings are consistent with prior reports highlighting the limitations of using ICD-based algorithms to identify COPD patients identified by spirometry (8–10). Given the relatively strong predictive power of the PRS_spiro in lung function defined COPD cohorts, rather than demonstrating that the PRS_spiro lacks utility, we believe the PRS_spiro may instead help identify high-risk individuals that do not have airflow obstruction (45).

The PRS_Asthma was newly developed in this study and predicted asthma, ACO, and asthma exacerbations, demonstrating the potential power of genetic variants, which have the added benefit of being available in early life and being antecedent to disease course or treatment. However, we acknowledge that the predictive power of the PRS_Asthma was low and we highlight other issues that preclude immediate clinical implementation. Compared to the asthma PRS developed by Sordillo et al. (22), we also used lassosum to develop a PRS, but 1) we used a larger GWAS from Han et al. (15) as opposed to Demenais et al. (13), 2) our PRS was tuned to an external validation population instead of a subset of participants in the target dataset, and 3) we systematically compared our score to a score derived from the larger, less well-phenotyped GBMI asthma GWAS. The PRS_Asthma derived from Han et al. (15) GWAS summary statistics utilized a smaller number of well-phenotyped asthma patients, yet had better predictive performance than a PRS derived from the GBMI GWAS (14) which included ~3-fold the number of participants. These results highlight the importance of identifying clinically-meaningful cases and the potential for imprecise classification when leveraging biobank data. Taken together, our results suggest that the PRS translation from research to clinical cohorts depends, at least partially, on the study phenotype and associated definitions.

Broadly, our findings also identify areas that require further development. The effect size heterogeneity in our PRS_spiro suggests the need to address the issue that patients identified as having COPD in real-world cohorts may not meet spirometric COPD criteria. Similar to comparisons of efficacy versus effectiveness in clinical trials, the performance of PRSs in real-life populations may be attenuated or require different approaches. In addition, our work highlights a need for improved asthma definitions, larger GWASs including asthma-specific imaging and other phenotypes, and the need for larger multi-ancestry genetic cohorts.

ACO individuals exhibited higher predicted genetic risk for both asthma and low spirometry measures compared to those with normal spirometry or moderate-to-severe COPD alone. ACO patients have been observed to have increased airway wall thickness (4), and we observed that higher PRS_Asthma scores were associated with higher Pi10 and WA% in COPDGene AA participants; higher PRS_spiro scores were also associated with higher Pi10 and WA % in this and prior studies (23,24). John et al. reported substantial genetic correlations between ACO and COPD (rg=0.83), ACO and asthma (rg=0.74), and ACO and FEV₁/FVC (rg= −0.69) (19). Overall, the findings that the PRS_spiro predicts asthma as well as ACO are consistent with shared genetic pathways of reduced lung function in COPD and asthma, and highlights the importance of increasing our understanding of the genetics of COPD and asthma with respect to disease classification.

Individuals with a prior history of asthma (diagnosed < 40 years of age) combined with a higher PRS_spiro were at greater risk of more severe COPD phenotypes (moderate-to-severe COPD and >5% emphysema) at COPDGene study enrollment. Whether these individuals progressed from asthma to COPD or were misdiagnosed early in life is unclear. The former explanation would lend evidence for the ‘Dutch hypothesis’ suggesting that asthma and COPD have common genetic origins and diverge into separate diseases (11,12). When interpreted from this perspective, our results suggest that disentangling the biological underpinnings of the PRS_spiro may allow early intervention and prevention of destructive emphysema in young asthmatics. From the perspective of asthma misdiagnosis, we observed that including PRSs in multivariable models with known clinical risk factors led to a net reclassification improvement of about 2%; while this improvement was statistically significant, the extent to which it translates into clinical benefits requires further investigation. Our results are also consistent with a recent report that the PRS_spiro can predict COPD occurring early in life (26) and extends this concept to individuals carrying only a diagnosis of asthma.

We found that both PRSs predict asthma exacerbations in GERA. Asthma exacerbations are complex traits that are traditionally challenging to predict because of multiple causal factors with small effect sizes, small sample sizes, and interactions amongst causal factors (46). Our results suggest that genetic factors that lead to the presence or absence of asthma and COPD can also predict severity. Whether a PRS could be developed specifically to predict asthma exacerbations and incorporated into a machine learning model to improve asthma exacerbation predictions requires further study.

The strengths of the current study include developing and comparing an asthma PRS derived from a smaller, well-phenotyped study, to one derived from a larger study with asthma phenotyping based entirely on administrative coding, testing PRSs in four large cohorts, both research-based and real-life, across four ancestry groups, and utilizing a novel approach of using two PRSs (PRS_Asthma and PRS_spiro) to dissect asthma and ACO heterogeneity. Nevertheless, it is important to highlight that we observed markedly reduced performance in non-European populations, which can ideally be addressed as new methodological approaches improve cross-ancestry genetic prediction. While including more non-Europeans in the training dataset could theoretically improve cross-ancestry portability, Wang, et al., who despite leveraging GBMI GWAS summary statistics for polygenic risk prediction, found that performance was best across traits using European LD reference panels, reflecting the fact that biobanks still include a majority of European-descent individuals (21). It is unclear if the PRS_Asthma represents genetic determinants of airway inflammation or hyperresponsiveness, response to infections, or other genetic factors associated with asthma. The biological underpinnings of both PRSs may be able to be dissected as new genomic methods emerge allowing integration of genetics with additional Omics data types. We defined ACO based on prior publications (4,5,19,47), which used both spirometry and self-report of asthma to define this phenotype. Ideally, we could have replicated the ACO findings in other cohorts, but spirometry was not available in GERA and MGB Biobank. CAMP participants were children with asthma, a subset of which developed reduced lung function in their twenties and thirties; we previously showed in this cohort that the PRS_spiro was associated with reduced lung function growth and decline trajectories (23), which had greater proportions of individuals meeting GOLD II or III COPD criteria early in life (29). However, it is important to note that these CAMP participants likely represent a distinct phenotype from the COPDGene ACO participants included in the present study. COPDGene participants were older smokers with COPD, a self-reported asthma diagnosis, and presumably had concomitant clinical features of asthma; this is the more usual context in which an ACO diagnosis is applied. Therefore, we did not consider CAMP participants with low lung function to have ACO, though future investigation into this issue could be undertaken. Finally, while ACO patients and certain patients with asthma diagnosed before age 40 may exhibit high risk for airflow obstruction based on their PRS_spiro, there remain few therapeutic options except for smoking cessation. However, there is some evidence to suggest that genetic knowledge of COPD risk can improve smoking cessation attempts (48).

In conclusion, PRSs for asthma and spirometry were both associated with ACO and asthma exacerbations. Further, the PRS_spiro could predict which individuals with asthma will develop COPD with emphysema. This study offers insight into how genetic heterogeneity correlates with phenotypic heterogeneity in asthma and ACO. Further research is needed to understand the mechanistic molecular underpinnings of our observations and the optimal approaches to translate genetic prediction tools into real-life cohorts.

Key Messages.

• We developed a polygenic risk score (PRS) for asthma (PRS_Asthma) and compared prediction using a previously-published spirometry based COPD PRS (PRS_spiro).

• The PRS_Asthma was associated with asthma, asthma-COPD overlap (ACO), and asthma exacerbations, while the PRS_spiro was associated with spirometry-defined COPD and ACO, but not electronic health record (EHR)-based COPD phenotypes.

• Polygenic risk scores for asthma and spirometry demonstrate promise in clinical utility, though discrepancies in genetic prediction performance in research versus EHR-based cohorts necessitates further investigation.

Abbreviations:

AA

African American

ACO

asthma-chronic obstructive pulmonary disease overlap

CAMP

Childhood Asthma Management Program study

COPD

chronic obstructive pulmonary disease

COPDGene

Genetic Epidemiology of COPD study

EHR

electronic health record

GBMI

Global Biobank Meta-analysis Initiative

GenKOLS

Genetics of COPD Norway study

GERA

Genetic Epidemiology Research on Adult Health and Aging cohort

GOLD

Global Initiative for Obstructive Lung Disease

GWAS

genome-wide association study

HU

Hounsfeld units

ICD

International Classification of Diseases

IgE

immunoglobulin E

KPNC

Kaiser Permanente Northern California

% LAA < −950 HU

quantitative emphysema on inspiratory CT scans

LASSO

least absolute shrinkage and selection operator

LD

linkage disequilibrium

MGB

Mass General Brigham

NHW

non-Hispanic white

OR

odds ratio

Perc15

15th percentile of lung density histogram on inspiratory CT scans

Pi10

square root of wall area of a hypothetical internal perimeter of 10 mm

PRS

polygenic risk score

SD

standard deviation

TAGC

Trans-National Asthma Genetic Consortium

WA%

airway wall area percent on computed tomography