Assessing the contribution of rare genetic variants to phenotypes of chronic obstructive pulmonary disease using whole-genome sequence data

Abstract

Rationale: Genetic variation has a substantial contribution to chronic obstructive pulmonary disease (COPD) and lung function measurements. Heritability estimates using genome-wide genotyping data can be biased if analyses do not appropriately account for the nonuniform distribution of genetic effects across the allele frequency and linkage disequilibrium (LD) spectrum. In addition, the contribution of rare variants has been unclear. Objectives: We sought to assess the heritability of COPD and lung function using whole-genome sequence data from the Trans-Omics for Precision Medicine program. Methods: Using the genome-based restricted maximum likelihood method, we partitioned the genome into bins based on minor allele frequency and LD scores and estimated heritability of COPD, FEV₁% predicted and FEV₁/FVC ratio in 11 051 European ancestry and 5853 African-American participants. Measurements and Main Results: In European ancestry participants, the estimated heritability of COPD, FEV₁% predicted and FEV₁/FVC ratio were 35.5%, 55.6% and 32.5%, of which 18.8%, 19.7%, 17.8% were from common variants, and 16.6%, 35.8%, and 14.6% were from rare variants. These estimates had wide confidence intervals, with common variants and some sets of rare variants showing a statistically significant contribution (P-value < 0.05). In African-Americans, common variant heritability was similar to European ancestry participants, but lower sample size precluded calculation of rare variant heritability. Conclusions: Our study provides updated and unbiased estimates of heritability for COPD and lung function, and suggests an important contribution of rare variants. Larger studies of more diverse ancestry will improve accuracy of these estimates.

Introduction

Chronic obstructive pulmonary disease (COPD), one of the leading causes of death worldwide, is diagnosed by a decrease in two key measurements of spirometry, forced expiratory volume in one second (FEV₁) and its ratio to forced vital capacity (FEV₁/FVC). Genetic factors are important risk factors for the development of COPD (1). Genome-wide association genomic studies (GWAS) for COPD and lung function measurements have identified hundreds of associated regions with common variants, with relevant effects in functional assays identified in genes such as IREB2, CHRNA3/5, HHIP, FAM13A, DSP, HTR4, LRP1 and CYP2A6 (2–10). Rare variants also affect COPD and related phenotypes as demonstrated in alpha-1 antitrypsin deficiency (1). Recent large-scale whole-genome sequencing (WGS) data from the National Heart, Lung and Blood Institute (NHLBI) Trans-Omics for Precision Medicine (TOPMed) program have enabled us to identify a subset of rare variants with putative associations to COPD and related phenotypes, including ARHGEF17 and CRISP1 (11).

Heritability is a measure that can provide relative estimates of the contribution of genetic versus environmental factors. Accurate determinations of heritability are important to determine the relative contribution of genetic variants, the potential performance of genetic risk scores and the contribution of rare variants. In prior twin studies and family studies, estimates of family-based heritability of COPD-related phenotypes ranged from 38% to 66% (12–14). Estimates in unrelated subjects can be obtained from genome-wide array data, and a recent study estimated heritability for COPD-related phenotypes using array data of ~35% (15). In another study using array data in the Framingham Heart Study (FHS), single nucleotide variant (SNV)-based heritability was estimated to be between 50% and 65% using SNVs with minor allele frequency (MAF) > 0.5% (14), with some of family-based heritability recovered from low-frequency SNVs (0.5% < MAF1%).

However, these previous studies have some limitations. First, heritability captured by rare variants has not been systematically assessed. Most low-frequency and rare variants are not captured by most genotyping arrays (16–18). In general, SNP arrays do not perform well for detecting, or imputing, rare variants (19,20). Determining the contribution of rare variants requires large subject sizes and WGS data. Second, prior estimates of heritability of lung function and COPD have used a single genetic relationship matrix (GRM). A GRM is calculated based on genetic variants to quantify genetic similarities between distantly related individuals. It has previously been shown that a single GRM approach can lead to biased estimates of heritability when causal variants have a different MAF or linkage disequilibrium (LD) properties than variants used in the analysis. Specifically, if causal variants are rarer (or more common) than variants used in the analysis, the estimate of heritability is biased downward (or upward) (21,22). This can be solved by using MAF-stratified GRMs in a model. In addition, if causal variants are enriched in genomic regions with lower (or higher) LD than average, the heritability estimate is downward (or upward) biased. Similar to the uneven MAF spectrum of causal variants, bias of heritability estimates can be solved by stratifying variants by their LD scores. To address these issues, Yang et al. (21) proposed a statistical method, termed the LD and MAF stratified genome-based restricted maximum likelihood (GREML-LDMS) approach, that creates bins of variants in different allele frequencies and LD thus reducing the bias of heritability estimates.

The recent advent of large-scale WGS data from the NHLBI TOPMed program enables an updated and more comprehensive assessment of the contribution of rare variants (11). Here we report SNV-based heritability of COPD and related phenotypes (FEV₁% predicted and FEV₁/FVC) using GREML-LDMS in nearly 12 000 unrelated individuals from high-coverage WGS data. Study participants were from four studies including three population-based studies and one COPD-enriched study as part of the TOPMed program. We also leveraged high-coverage WGS data to assess the proportion of phenotypic variance of COPD and related phenotypes explained by rare variants.

Results

Heritability estimates using a single GRM

We first calculated heritability using the standard method of a single component genetic relationship matrix (GREML-SC) (23). We estimated GRMs for rare variants (MAF < 0.01) and common variants (MAF0.01), then conducted GREML-SC analysis using each GRM. In European ancestry participants (N = 11 501), for variants with MAF < 0.01, none of heritability estimates were statistically significantly >0 (Table 1). For variants with MAF0.01, we estimated heritability to be 23.3% [standard error (SE) = 3.6%, P-value = 4.510⁻¹¹] for FEV₁% predicted, 17.9% (SE = 3.2%, P-value = 8.710⁻⁹) for FEV₁/FVC and 18% (SE = 5%, P-value = 1.610⁻⁴) for COPD. The heritability estimates of height and body mass index (BMI) were 54.8% (SE = 3.6%, P-value = 8.210⁻⁵²) and 21.1% (SE = 3.6%, P-value = 1.510⁻⁹), respectively.

Table 1.

Heritability estimates and SE using GREML-SC in European ancestry

Phenotype	Estimates		Rare variants	Common variants
FEV1 % predicted	(SE)		0.031 (0.041)	0.233 (0.036)
	95% CI	One-sided	(−0.037,)	(0.173,  )
		Two-sided	(−0.05, 0.112)	(0.162, 0.303)
	P-value		0.225	4.47  10  –11
FEV1/FVC	(SE)		−0.003 (0.032)	0.179 (0.032)
	95% CI	One-sided	(−0.057, )	(0.126,  )
		Two-sided	(−0.067, 0.06)	(0.117, 0.241)
	P-value		0.54	8.74  10  −9
COPD	(SE)		−0.141 (0.029)	0.18 (0.05)
	95% CI	One-sided	(−0.189, )	(0.098,  )
		Two-sided	(−0.198, −0.083)	(0.082, 0.278)
	P-value		1	1.55  10  −4
Height	(SE)		0.071 (0.044)	0.548 (0.036)
	95% CI	One-sided	(−0.001, )	(0.488,  )
		Two-sided	(−0.015, 0.157)	(0.477, 0.619)
	P-value		0.052	8.20  10  −52
BMI	(SE)		−0.026 (0.036)	0.211 (0.036)
	95% CI	One-sided	(−0.086, )	(0.152,  )
		Two-sided	(−0.097, 0.046)	(0.141, 0.281)
	P-value		0.758	1.52  10  −9

Estimates for FEV₁% predicted and FEV₁/FVC were adjusted for ascertainment bias. Rare and common variants stand for variants with MAF < 1% and MAF1%, respectively. Bold indicates statistically significant under the significance level of 0.05. One-sided 95% CI means 95% confidence interval for the alternative hypothesis, . Two-sided 95% CI means 95% confidence interval for the alternative hypothesis, .

In African-American participants (N = 5853), for common variants, the estimated heritability was 21.6% (SE = 7.8%, P-value = 0.003) for FEV₁% predicted, 25.3% (SE = 7.1%, P-value = 1.810⁻⁴) for FEV₁/FVC, 69% (SE = 7.4%, P-value = 7.910⁻²¹) for height and 27.7% (SE = 7.8%, P-value = 2.110⁻⁴) for BMI (Supplementary Material, Table S1). Heritability estimation for COPD, and for rare variant analysis of height and FEV₁/FVC, did not converge because of smaller sample size among African ancestry participants; thus, subsequent analyses using the GREML-LDMS methods (21) were restricted to European Ancestry.

Aggregated heritability estimates using WGS in European ancestry participants

Multiple studies have demonstrated a reduction in bias when variants are partitioned by MAF and degree of LD (21,22,24). Thus, we stratified all variants by eight bins based on their MAF in the dataset (0MAF < 0.0001, 0.0001MAF < 0.001, 0.001MAF < 0.01, 0.01MAF < 0.1, 0.1MAF < 0.2, 0.2MAF < 0.3, 0.3MAF < 0.4, 0.4MAF 0.5). Each of the eight MAF bins were further stratified into a low-LD bin and a high-LD bin based on the median value of LD scores of each MAF bin. Finally, we generated GRMs for 16 variant bins and performed GREML-LDMS using GCTA software. To get heritability estimates of rare and common variants, we aggregated heritability estimates for GRMs with MAF < 0.01 and MAF0.01, respectively.

The aggregated estimates of SNV-based heritability using GREML-LDMS are displayed in Table 2 and Supplementary Material, Table S2. To first confirm that our approach in this dataset yielded similar estimates to prior reports, we applied GREML-LDMS to height and BMI, and estimated a heritability of 82.4% (SE = 16.7%, P-value = 3.910⁻⁷) for height and 51.7% (SE = 15.3%, P-value = 3.510⁻⁴) for BMI. These estimates are slightly higher than the previously reported estimates (56–79% for height and 22–40% for BMI, Supplementary Material, Table S3) (21,22,24) that did not include the rarest variant bin.

Table 2.

Aggregated heritability estimates and SE for FEV₁% predicted, FEV₁/FVC and COPD using GREML-LDMS in European ancestry

Phenotype	LD	Estimates		Rare variants	Common variants	Total
FEV1 % predicted	High-LD	(SE)		0.05 (0.041)	0.075 (0.023)	0.125 (0.046)
		95% CId	One-sided	(−0.018, )	(0.037,  )	(0.049,  )
			Two-sided	(−0.03, 0.13)	(0.029, 0.121)	(0.035, 0.215)
		P-value		0.111	6.62  10  −4	0.003
	Low-LD	(SE)		0.308 (0.152)	0.122 (0.061)	0.43 (0.152)
		95% CI	One-sided	(0.058,  )	(0.021,  )	(0.18,  )
			Two-sided	(0.011, 0.605)	(0.002, 0.242)	(0.133, 0.727)
		P-value		0.021	0.023	0.002
	Total	(SE)		0.358 (0.174)	0.197 (0.06)	0.556 (0.172)
		95% CI	One-sided	(0.071,  )	(0.098,  )	(0.272,  )
			Two-sided	(0.017, 0.699)	(0.079, 0.316)	(0.219, 0.892)
		P-value		0.02	5.28  10  −4	6.06  10  −4
FEV1/FVC	High-LD	(SE)		0.046 (0.035)	0.06 (0.022)	0.105 (0.04)
		95% CI	One-sided	(−0.012, )	(0.023,  )	(0.039,  )
			Two-sided	(−0.023, 0.115)	(0.016, 0.103)	(0.026, 0.185)
		P-value		0.097	0.003	0.005
	Low-LD	(SE)		0.101 (0.142)	0.119 (0.058)	0.219 (0.142)
		95% CI	One-sided	(−0.133, )	(0.023,  )	(−0.015, )
			Two-sided	(−0.177, 0.379)	(0.005, 0.232)	(−0.059, 0.498)
		P-value		0.239	0.02	0.061
	Total	(SE)		0.146 (0.162)	0.178 (0.057)	0.325 (0.16)
		95% CI	One-sided	(−0.121, )	(0.084,  )	(0.061,  )
			Two-sided	(−0.171, 0.464)	(0.067, 0.29)	(0.012, 0.638)
		P-value		0.183	8.54  10  −4	0.021
COPD	High-LD	(SE)		0.073 (0.042)	0.024 (0.033)	0.098 (0.052)
		95% CI	One-sided	(0.003,  )	(−0.03, )	(0.012,  )
			Two-sided	(−0.01, 0.156)	(−0.04, 0.088)	(−0.004, 0.199)
		P-value		0.042	0.229	0.03
	Low-LD	(SE)		0.093 (0.187)	0.164 (0.087)	0.257 (0.188)
		95% CI	One-sided	(−0.215, )	(0.02,  )	(−0.053, )
			Two-sided	(−0.273, 0.459)	(−0.007, 0.336)	(−0.112, 0.626)
		P-value		0.309	0.03	0.086
	Total	(SE)		0.166 (0.211)	0.188 (0.086)	0.355 (0.208)
		95% CI	One-sided	(−0.182, )	(0.047,  )	(0.011,  )
			Two-sided	(−0.248, 0.581)	(0.02, 0.357)	(−0.054, 0.763)
		P-value		0.215	0.014	0.044

Estimates for FEV₁% predicted, FEV₁/FVC were adjusted for ascertainment bias. Heritability estimates were aggregated by MAF (rare variants and common variants) and LD (high-LD versus low-LD) groups. Rare and common variants stand for variants with MAF < 1% and MAF1%, respectively. Bold indicates statistically significant under the significance level of 0.05. One-sided 95% CI means 95% confidence interval for the alternative hypothesis, . Two-sided 95% CI means 95% confidence interval for the alternative hypothesis, .

For lung function measurements, after adjusting for ascertainment bias, we estimated the heritability of FEV₁% predicted to be 55.6% (SE = 17.2%, P-value = 6.110⁻⁴), and rare variants and common variants accounted for 35.8% (SE = 17.4%, P-value = 0.02) and 19.8% (SE = 6%, P-value = 5.310⁻⁴), respectively. The largest contributor to heritability was from rare variants in the low-LD group, though with large SEs (h² = 30.8%, SE = 15.2%, P-value = 0.021). When we aggregated the heritability by LD groups, 43% (SE = 15.2%, P-value = 0.002) and 12.5% (SE = 4.6%, P-value = 0.003) of heritability were accounted for by variants in the low-LD group and high-LD group, respectively.

The overall heritability of FEV₁/FVC was estimated to be 32.5% (SE = 16%, P-value = 0.021). A total of 14.6% (SE = 16.2%) and 17.8% (SE = 5.7%, P-value = 8.510⁻⁴) of the heritability were contributed by rare variants and common variants, respectively. In terms of LD structure, variants in low-LD group and high-LD group explained 21.9% (SE = 14.2%) and 10.5% (SE = 4%, P-value = 0.005) of the heritability, respectively.

For COPD, the heritability estimate was 35.5% (SE = 20.8%, P-value = 0.044) and of those, 16.6% (SE = 21.1%) and 18.8% (SE = 8.6%, P-value = 0.014) were explained by rare variants and common variants, respectively. Similar to the lung function measurements, heritability was mostly contributed by variants in the low-LD group (low-LD: h² = 25.7%, SE = 18.8%, P-value = 0.086; high-LD: h² = 9.8%, SE = 5.2%, P-value = 0.03). These estimates showed higher SEs compared with lung function measurements due in part to a reduction in sample size from excluding participants who were neither cases nor controls.

Significant heritability estimates of individual MAF and LD bins in European ancestry participants

The heritability estimates for individual MAF and LD bins are shown in Supplementary Material, Table S4 and Figure 1. For FEV₁% predicted, variants with 0.001MAF < 0.1 showed the largest contribution to the phenotype among MAF bins (h² = 23.8%, SE = 13%, P-value = 0.033). In the high-LD group, heritability estimates of the variants with 0.2MAF < 0.3 and 0.4MAF0.5 were significantly larger than zero (h² = 2.4%, SE = 1.1%, P-value = 0.017 and h² = 1.9%, SE = 1%, P-value = 0.023, respectively). In the low-LD group, heritability estimates of the variants with 0.3MAF < 0.4 was statistically significant (h² = 6.8%, SE = 2.8%, P-value = 0.008). For FEV₁/FVC, we found two significant heritability estimates of the variants with 0.2MAF < 0.3 in the high-LD group (h² = 2.5%, SE = 1.1%, P-value = 0.013) and 0.3MAF < 0.4 in the low-LD group (h² = 5.7%, SE = 2.7%, P-value = 0.016). For COPD disease status, a substantial fraction of heritability was captured by the rarest MAF bin (0MAF < 0.0001) with estimates of 32.6% (SE = 13.6%, P-value = 0.008), and heritability estimates for its LD-stratified bins were also statistically significant (high-LD: h² = 6.4%, SE = 3.2%, P-value = 0.021; low-LD: h² = 26.2%, SE = 11%, P-value = 0.008). We observed significant heritability estimates of 9.7% among rare variants (0.001MAF < 0.01) in the high-LD group (SE = 4.1%, P-value = 0.009). Compared with the heritability estimates using GREML-SC, the SC method did not capture heritability of rare variants.

Figure 1.

Heritability estimates stratified in 16 bins (eight MAF bins and two LD bins) in European ancestry. The bars display SEs.

Discussion

A genetic contribution to COPD and lung function has been recognized for decades, based on family-based and twin studies and more recently, from genome-wide genotyping data, widely available in population-based samples. However, these methods relied on the assumptions of a normal distribution of genetic variant effect sizes independent of LD and inversely proportional to the MAF. In addition, despite examples of rare variant contributions to complex disease, many studies have not been able to identify and replicate rare variants of statistically significant effect (25–27). Prior studies in COPD and lung function have either not identified statistically significant variants, or identified variants that have not been consistently replicated (28–33).

Our study sought to use comprehensive WGS data available in four cohorts to provide unbiased estimates of heritability and to examine the contribution of rare variants on COPD and lung function. Compared with prior estimates in a COPD-enriched study (15), comprised of smokers with and without COPD, our study of population-based and COPD-enriched participants resulted in a smaller heritability estimate from common variants than the previous study. Compared with a prior study of low-frequency variants (0.5%MAF < 5%) in the UK Biobank, our heritability estimates of rare variant (MAF < 1%) for FEV₁/FVC were larger (34) suggesting nonzero contribution of rare variants to these phenotypes. Our heritability estimates of common variants based on GREML-LDMS and GREML-SC indicate that bias of common variants because of heritability partitioning was negligible. However, we identified substantial differences in the estimates of rare variants depending on stratification of GRMs. Although the estimates based on the GREML-SC approach were nearly zero, 35.8%, 14.6% and 16.6% of heritability of rare variants were recovered for FEV₁% predicted, FEV₁/FVC and COPD by using GREML-LDMS, respectively. Therefore, rare causal variants may lie within specific MAF or LD bins, whereas common causal variants may spread among multiple MAF and LD bins. Another possible explanation is that the genetic similarity between individuals modeled by different common variant bins are more similar to each other than when using rare variant bins. This finding may be because of chance or because of a selection process that we do not model.

For all phenotypes, heritability estimates were predominantly explained by rare variants in low-LD bins with some, but not all, statistically significant estimates of heritability. These findings are consistent with other studies demonstrating that, in general, rare variants with lower levels of LD have higher levels of heritability (35). These variants also are more enriched in regions of functional annotation. We note that the majority of identified variants in this study, as in other sequencing studies are rare. These findings are consistent with the effects of selection in reducing deleterious alleles over time (35,36).

Heritability is defined between 0 and 1 but its estimate can exceed these values, following the normal distribution. Thus, to get an unbiased estimate, we allowed the heritability estimate to be negative. Indeed, we observed negative estimates with wide confidence interval in several individual MAF and LD bins. We kept negative estimates to obtain unbiased estimates of total heritability, although negative heritability has no meaning in itself and should be interpreted as zero heritability. In practice, we are likely to observe a negative estimate in two situations. First, when the sample size is small, the estimate has a chance to be out of the parameter space because of a large SE. Second, when the true heritability is very small, then we also have a certain probability to see negative estimate even though the sample size is large. Although our sample sizes were large for WGS data for COPD and lung function, the SEs of our estimates are large, resulting in only nominally (P-value < 0.05) significant P-values unadjusted for multiple testing, indicating that even larger sample sizes are needed to accurately identify the contribution of rare variants generally and also to determine the relative contribution of specific rare functional regions. Nevertheless, the estimates of the overall heritability were reliable. In our study, heritability estimation using GREML-LDMS was limited to European ancestry because heritability estimates in African ancestry were imprecise or did not converge because of sample size. Estimating and interpreting SNV-heritability is challenging in diverse populations (37) and GCTA-GREML that is the method estimating heritability from individual-level genetic data can generate biased estimates in the presence of population stratification (38). Our study thus further emphasizes the need to increase the number of participants of non-European ancestry. Finally, heritability is a measure that is specific to a population and set of environmental conditions. We studied lung function and disease status phenotypes available in our sample. In addition, the relative contribution of rare variants to other COPD-related phenotypes, such as those obtained from imaging studies, is unknown.

Overall, our results provide an updated estimate of heritability, in the largest cohort for lung function and COPD to undergo WGS to date. We find evidence for a contribution of rare variants to lung function and COPD and an enrichment in low-LD regions. These findings support the ongoing investigation of rare variants in COPD.

Materials and Methods

Study participants

We selected 11 849 European ancestry participants and 7167 African-American participants from four population-based cohorts [the Atherosclerosis Risk in Communities (ARIC) Study, the FHS, the Multi-Ethnic Study of Atherosclerosis (MESA) and Jackson Heart Study (JHS)] and one COPD-enriched study [the Genetic Epidemiology of COPD (COPDGene) Study] as a part of NHLBI TOPMed program. The ancestry was defined by genetic data. The study descriptions for each cohort are provided in the Supplementary Material, Text S2.

Genetic variants were extracted from the Freeze 5b version of WGS data in the TOPMed program (39). Details on sequencing method of TOPMed are found at https://www.nhlbiwgs.org/topmed-whole-genome-sequencing-project-freeze-5b-phases-1-and-2. Briefly, joint-subject variant calling identified ~ 470M genetic variants for 54 499 participants which is a subset of TOPMed participants on human genome assembly GRCh38 with deep coverage (~30×). Among these variants, we only considered genetic variants carried by at least one participant in the cohorts included in this analysis.

Quality control

The TOPMed callset includes extensive quality control, detailed at https://www.nhlbiwgs.org/topmed-whole-genome-sequencing-project-freeze-5b-phases-1-and-2. We additionally excluded SNVs with missingness rate > 5% in our subsample, or a P-value of the Hardy–Weinberg equilibrium test < 110⁻⁵. We considered only biallelic SNVs on the autosomes. We also excluded any participants whose genotype missingness rate was >5%. To include only unrelated participants, we estimated kinship coefficients of every pair of participants using PC-Relate (40) and we randomly excluded one of each pair of participants with estimates of kinship coefficient > 0.05. After filtering, we retained 11 501 European ancestry participants and 5853 African ancestry participants with ~123M and ~112M SNVs, respectively (Table 3 and Figure 2). We confirmed that there were no related participants on the reported pedigree after quality control of the samples. All participants that were not European-ancestry or African-American were excluded from the analysis because of sample size. As expected, the COPD-enriched study (COPDGene) showed lower mean values of lung function and higher pack-years than the population-based studies in both ancestries.

Table 3.

Descriptive statistics of study participants after quality control

Phenotype	Ancestry		Study
	European	African	COPD-enriched	Population-based
Number of subjects	11 051	5853	8397	8507
COPD subjects	3234	930	3120	1044
Gender (F/M)	5460/5591	3025/2828	3907/4490	4578/3929
Age	61.64 (9.65)	57.61 (10.43)	59.5 (9.08)	60.98 (10.99)
Pack-years	32.04 (29.58)	22.28 (24.47)	44.36 (25)	13.17 (22.16)
Height	169.03 (9.54)	169.14 (9.8)	170.2 (9.47)	167.95 (9.65)
BMI	28.18 (5.63)	30.07 (6.69)	28.88 (6.26)	28.79 (5.91)
FEV1 % predicted	80.24 (24.55)	86.13 (22.27)	72.86 (25.58)	91.57 (17.88)
FEV1/FVC	0.68 (0.14)	0.75 (0.12)	0.65 (0.16)	0.75 (0.09)

We counted the number of subjects, COPD subjects and each gender. For continuous variables, mean and standard deviation were shown.

Figure 2.

Quality control for the SNVs and participants. Multiple standard quality controls were performed for the raw data to exclude outlier SNVs and participants.

Phenotypes

We estimated the heritability of spirometry measures, COPD affection status, as well as anthropometric values (the latter as ‘positive controls’). For lung function, we used percentage of predicted value of FEV₁ (FEV₁% predicted) and FEV₁/FVC ratio as continuous phenotypes. To ensure harmonized phenotypes across multiple TOPMed population studies, we used phenotypes derived from the protocol of the NHLBI Pooled Cohorts Study (Supplementary Material, Text S3) (41). We calculated percent predicted values of FEV₁ using Hankinson’s reference equations (42). We also examined the heritability of height and BMI within our participants, as these measures have been estimated previously in larger samples. Similar to a previous approach (24), to adjust for age, we regressed all continuous phenotypes on age within each sex and study, and residuals were standardized by sex-specific standard deviations. We then pooled the standardized residuals together and applied a rank-based inverse normal transformation to the standardized residuals. For the dichotomous phenotype, we defined COPD status of each participant using GOLD criteria for moderate-to-severe obstruction (43) as follows: (i) cases: FEV₁% predicted <80% and FEV₁/FVC < 0.7, (ii) controls: FEV₁% predicted 80% and FEV₁/FVC0.7. In addition to age and sex, we included pack-years of smoking as a covariate for all phenotypes except for height, as well as sequencing center, and first 10 principal components of genetic ancestry scores as fixed effects.

Statistical methods for estimating heritability

The overview of the workflow for the statistical analysis is illustrated in Figure 3. GREML, which is implemented in the GCTA software, is a statistical method to estimate genetic and environmental variance components using a linear mixed model (21,23). For a quantitative trait , the model of GREML is given by

where is a vector of coefficients of the covariates such as age, sex or principal component scores, is a random effect of ^th SNV-set with a corresponding GRM , which follows and is a vector of residual effects following . For the ^th GRM, , the genetic relationship between individuals and was calculated as follows,

where is the number of SNVs, is a minor allele count of ^th SNV for individual and is an MAF of ^th SNV in the dataset. The variance–covariance matrix of a quantitative trait is and we estimate variance components, and , using the restricted maximum likelihood approach (44). The proportion of phenotypic variance explained by the ^th set of SNV and by all the SNVs are defined as and , respectively.

Figure 3.

Workflow of statistical analysis for the heritability estimation of COPD-related phenotypes. Overview of the statistical analysis to estimate heritability of the phenotypes is illustrated.

Because our study includes a COPD-enriched population, study participants had lower values of FEV₁% predicted or FEV₁/FVC than the general population on average. To estimate heritability of FEV₁% predicted and FEV₁/FVC in the general population, we adjusted the heritability estimates in the ascertained participants by using their relationship expressed as a function of proportions of participants at the specific threshold value of the phenotype in both the general population and the ascertained participants, and the population mean and variance of the phenotype. To adjustment ascertainment bias, we expressed heritability in the participants as a function of the heritability in the population given that the phenotype and genetic factor are jointly normally distributed; the heritability in the population can be obtained by inverting the equation. The detailed derivation is provided in the Supplementary Material, Text S4.

Similarly, heritability for a dichotomous phenotype, COPD status in our case, needs to be adjusted when cases and controls are not a random sample from the general population (45). Here, we assume that cases and controls are determined by the unobserved continuous liability score and the threshold . Under the assumption, participants are considered as a case if their liability scores are larger than the threshold ; otherwise, they are considered as a control. The threshold is chosen to maintain the assumed prevalence q. In the absence of case–control ascertainment, the heritability of a dichotomous trait on the liability scale, , can be expressed with the heritability on the observed 0–1 scale, , as follows:

where the population prevalence and the value of the probability density function for the standard normal distribution at the threshold

Under the liability threshold model, the heritability of a dichotomous phenotype on the liability scale, , can be expressed with the heritability on the observed 0–1 scale, , as follows:

where the population prevalence, the proportion of the cases in the sample and the value of the probability density function for the standard normal distribution at the threshold . We assumed a 10% prevalence of COPD in the population and the same prevalence was assumed for all studies (46).

Partitioning heritability

To assess the heritability of COPD and lung function using methods comparable to prior studies and to serve as a baseline to examine the effects of partitioning on heritability, we estimated heritability using single-component GREML analysis (GREML-SC) in GCTA software for common variants (MAF0.01) and rare variants (MAF < 0.01) with the same fixed effects in the GREML-LDMS analysis.

According to previous studies demonstrating a reduction in bias when variants are partitioned by allele frequency and degree of LD (21,22,24), we stratified all variants by eight bins based on their MAF in the dataset (0MAF < 0.0001, 0.0001MAF < 0.001, 0.001MAF < 0.01, 0.01MAF < 0.1, 0.1MAF < 0.2, 0.2MAF < 0.3, 0.3MAF  < 0.4, 0.4MAF0.5). We then calculated LD score of the variants on a sliding window of 10 Mb within each MAF bin as follows:

where is a mean between the target variant and all other variants in the window and is the number of variants in the window. Here, LD score is defined as the sum of between a variant and all the variants in a window. We used the GCTA software for the LD score calculation (23). On each chromosome, we further stratified each of the eight MAF bins into a low-LD bin and a high-LD bin based on the median value of LD scores of each MAF bin, and combined all variants in the same MAF and LD bin across chromosomes. The actual median LD scores in MAF bins were provided in Supplementary Material, Table S5. In total, we generated 16 GRMs and performed GREML-LDMS using GCTA software. To estimate aggregated heritabilities of rare and common variants, we summed up all heritability estimates for GRMs with MAF < 0.01 and MAF0.01, respectively, and its SE was approximately derived based on the Delta method (Supplementary Material, Text S5).

Among ~ 123M SNVs in 11 502 unrelated European ancestry participants, 93.53% were rare variants (MAF < 0.01) and 6.47% were common variants (MAF0.01). SNVs with MAF < 0.001 consisted of 89.38% of all variants. The remaining MAF bins (0.001MAF < 0.01, 0.01MAF < 0.1, 0.1MAF < 0.2, 0.2MAF < 0.3, 0.3MAF  < 0.4 and 0.4MAF0.5) each accounted for <5% of total SNVs, with a decreasing proportion of total SNVs with increasing MAF (Fig. 4).

Figure 4.

MAF distribution of variants in WGS data. Values indicate MAF and the proportion of variants in the MAF groups (%).

Code Availability

The analysis is readily applicable in other diseases. The code used in the study is provided at https://github.com/wonjikim11/RV_h2_COPD. Detailed information and additional options for GCTA-GREML analysis can be found in the GCTA online manual. (https://yanglab.westlake.edu.cn/software/gcta/#Overview).

Funding

Whole-genome sequencing (WGS) for the Trans-Omics in Precision Medicine (TOPMed) program was supported by the National Heart, Lung and Blood Institute (NHLBI). WGS for `NHLBI TOPMed: Atherosclerosis Risk in Communities (ARIC)' (phs001211) was performed at the Baylor College of Medicine Human Genome Sequencing Center (HHSN268201500015C and 3U54HG003273-12S2) and the Broad Institute for MIT and Harvard (3R01HL092577-06S1). WGS for `NHLBI TOPMed: Whole Genome Sequencing and Related Phenotypes in the Framingham Heart Study (FHS)' (phs000974) was performed at the Broad Institute of MIT and Harvard (HHSN268201500014C and 3R01HL092577-06S1). WGS for `NHLBI TOPMed: Multi-Ethnic Study of Atherosclerosis (MESA)' (phs001416) was performed at the Broad Institute of MIT and Harvard (3U54HG003067-13S1 and HHSN268201500014C). WGS for `NHLBI TOPMed: Genetic Epidemiology of COPD (COPDGene) in the TOPMed Program' (phs000951) was performed at the University of Washington Northwest Genomics Center (3R01 HL089856-08S1) and the Broad Institute of MIT and Harvard (HHSN268201500014C). WGS for `NHLBI TOPMed: The Jackson Heart Study (JHS)' (phs000964) was performed at the University of Washington Northwest Genomics Center (HHSN268201100037C). Centralized read mapping and genotype calling, along with variant quality metrics and filtering were provided by the TOPMed Informatics Research Center (3R01HL-117626-02S1; contract HHSN268201800002I). Phenotype harmonization, data management, sample-identity QC and general study coordination, were provided by the TOPMed Data Coordinating Center (3R01HL-120393-02S1; contract HHSN268201800001I). Phenotype harmonization for pulmonary traits was contributed by the NHLBI Pooled Cohorts Study with funding from NIH/NHLBI R21 HL121457, R21 HL129924, K23 HL130627, R01 HL077612. This research was also supported by R01 HL149861, R01 HL135142, R01 HL137927 (M.H.C.) and R01 HL089856 and R01 HL147148 (M.H.C. and E.K.S.), and K01 HL129039 (D.Q.). The COPDGene study (NCT00608764) is also supported by NHLBI grants U01 HL089897 and U01 HL089856, and the COPD Foundation through contributions made to an Industry Advisory Board that has included AstraZeneca, Bayer Pharmaceuticals, Boehringer-Ingelheim, Genentech, GlaxoSmithKline, Novartis, Pfizer, and Sunovion.

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