Proteomic Risk Score of Increased Respiratory Susceptibility: A Multicohort Study

Abstract

Rationale

Accelerated decline in lung function is associated with incident chronic obstructive pulmonary disease (COPD), hospitalization, and death. However, identifying this trajectory with longitudinal spirometry measurements is challenging in clinical practice.

Objectives

To determine whether a proteomic risk score trained on accelerated decline in lung function can assess the risk of future respiratory disease and mortality.

Methods

In the Coronary Artery Risk Development in Young Adults Study, a population-based cohort starting in young adulthood, longitudinal measurements of FEV₁ percent predicted (up to six time points over 30 yr) were used to identify accelerated and normal decline trajectories. Protein aptamers associated with an accelerated decline trajectory were identified with multivariable logistic regression followed by LASSO (least absolute shrinkage and selection operator) regression. The proteomic respiratory susceptibility score was derived on the basis of these circulating proteins and applied to the U.K. Biobank (UKBB) and COPDGene studies to examine associations with future respiratory morbidity and mortality.

Measurements and Main Results

Higher susceptibility score was independently associated with all-cause mortality (UKBB hazard ratio [HR], 1.56; 95% confidence interval [CI], 1.50–1.61; COPDGene HR, 1.75 95% CI, 1.63–1.88), respiratory mortality (UKBB HR, 2.39; 95% CI, 2.16–2.64; COPDGene HR, 1.81; 95% CI, 1.32–2.47), incident COPD (UKBB HR, 1.84; 95% CI, 1.71–1.98), incident respiratory exacerbation (COPDGene odds ratio, 1.10; 95% CI, 1.02–1.19), and incident exacerbation requiring hospitalization (COPDGene OR, 1.17; 95% CI, 1.08–1.27).

Conclusions

A proteomic signature of increased respiratory susceptibility identifies people at risk of respiratory death, incident COPD, and respiratory exacerbations. This susceptibility score is composed of proteins with well-known and novel associations with lung health and holds promise for the early detection of lung disease without requiring years of spirometry measurements.

At a Glance Commentary

Scientific Knowledge on the Subject

Accelerated decline in lung function is a key feature of impaired respiratory health that reflects susceptibility to a variety of pulmonary risk factors and is associated with respiratory morbidity and mortality. However, repeated lung function measurements over a life course are often not performed in clinical practice. Identifying blood biomarkers that capture respiratory susceptibility at a single time point has real-world advantages. Although numerous studies have identified biomarkers associated with FEV₁ decline and risk of chronic obstructive pulmonary disease exacerbations, there have been no prior studies using large-scale proteomics to derive a blood-based biomarker of respiratory susceptibility defined by adult life course FEV₁ trajectory that further tested its associations with future respiratory disease and mortality.

What This Study Adds to the Field

Taking advantage of previously validated adult life course FEV₁ trajectories based on repeated lung function measures performed over 30 years in a general U.S. population cohort (CARDIA), we derived a proteomic signature of respiratory susceptibility trained on accelerated decline trajectory. The resulting proteomic susceptibility score was associated with all-cause mortality, respiratory mortality, incident chronic obstructive pulmonary disease, and respiratory exacerbations in two diverse cohorts. This susceptibility score holds the potential for the early detection of lung disease without requiring years of spirometry measurements.

In 2019, chronic respiratory diseases affected 454.6 million people and were responsible for 4.0 million deaths, making them the third leading cause of death worldwide (1). Chronic obstructive pulmonary disease (COPD) is the most common chronic respiratory disease and accounted for 3.3 million of these deaths (1). Although COPD is only diagnosed once pathophysiological changes have led to irreversible airflow limitation, it is now recognized that in many individuals, disease diagnosis is preceded by a life course of accelerated decline (AD) in lung function. Generating tools to identify these susceptible individuals paves the way for the development of preventive measures and disease-modifying treatments.

Fletcher and Peto first described that a major pathway to death and disability caused by COPD arises from an AD in FEV₁ (2). AD FEV₁ trajectory has since been associated with increased risk of emphysema, COPD-related hospitalization, and all-cause mortality (3–7). Indeed, loss of FEV₁ over the adult life course reflects susceptibility to a variety of pulmonary risk factors and subsequent host responses (8, 9). However, obtaining repeated measures of pulmonary function is seldom done in clinical practice. This is particularly true in a healthy general population, where there is debate whether spirometry may even be useful (10). Identifying individuals with increased risk of respiratory disease using a single time point measure rather than years of observation therefore has real-world advantages.

In this context, we identified circulating proteins associated with AD in FEV₁ trajectory to derive a blood-based biomarker signature that would identify increased respiratory susceptibility through a blood test at a single time point. Proteomics, a large-scale sampling of proteins (in this case from plasma), provides a unique opportunity to molecularly identify individuals susceptible to lung disease, especially in the absence of longitudinal lung function data. Here, we derived a proteomic risk score of increased respiratory susceptibility in the CARDIA (Coronary Artery Risk Development in Young Adults Study) cohort, a population-based cohort study with up to six lung function measurements over 30 years. We then tested the clinical relevance of this proteomic risk score in a separate population-based cohort, the U.K. Biobank (UKBB; 35,000 participants), and in an at-risk cohort of current and former smokers, the COPDGene (Genetic Epidemiology of COPD) study (5,168 participants). We hypothesized that our proteomic respiratory susceptibility score would effectively predict incident COPD, future respiratory exacerbations, and mortality. Some of the results of these studies were previously reported in the form of an abstract (11).

Methods

Study Design

This is a prospective cohort study evaluating the association between a proteomic risk score of increased respiratory susceptibility and future respiratory disease, morbidity, and mortality. The susceptibility score was derived in the CARDIA cohort and applied in the UKBB and COPDGene studies. The study design and timeline are depicted in Figure 1. CARDIA and COPDGene are reviewed annually by institutional review boards at each center, and UKBB is monitored by an independent ethics and governance council. All participants provided written informed consent.

Figure 1.

Study design and timeline. Top: In the CARDIA cohort, we identified participants with normal decline (n = 2,332) and accelerated decline (n = 138) in lung function trajectories. These trajectories were adapted from previously developed and published trajectories (7) based on up to six measurements of FEV₁ percent predicted taken at Years 0, 2, 5, 10, 20, and 30 (mean ages, 25, 27, 30, 35, 45, and 55 yr). Plasma samples drawn at Year 25 (mean age, 50 yr) were analyzed for large-scale protein identification (proteomics). We then examined which circulating proteins were associated with an accelerated decline in FEV₁ trajectory, independent of cross-sectional FEV₁ at Year 20. We used these proteins to derive the proteomic “respiratory susceptibility score.” Bottom: We applied the respiratory susceptibility score to the U.K. Biobank and COPDGene cohort studies. Plasma samples drawn at the baseline visit in U.K. Biobank and the Year 5 visit (visit 2) in COPDGene were analyzed for the proteins in the susceptibility score. We examined the association between susceptibility score and incident chronic obstructive pulmonary disease, respiratory exacerbations, respiratory death, and all-cause death. The median follow-up time for all-cause death was 13.7 years (interquartile range, 12.0–14.5) in U.K. Biobank and 6.5 years (interquartile range, 4.9–7.5) in COPDGene. Figure created with BioRender.com.

Study Populations

The CARDIA study is a population-based cohort study of 5,115 adults enrolled at ages 18–30 years from four centers in the United States. The susceptibility score was generated from and tested in 2,470 participants with complete data available at the Year 25 follow-up visit (see Figure E1 in the online supplement).

UKBB is a population-based cohort study of 502,625 adults between 40 and 69 years old from the United Kingdom, of whom 54,306 had some circulating proteomic data at the time of initial assessment. The proteomic susceptibility score was calculated for the 35,000 participants with complete proteomic data at the baseline visit.

COPDGene is a prospective study that represents a downstream at-risk cohort of 10,198 former and current smokers with at least a 10–pack-year smoking history, aged 45 to 80 years. The proteomic susceptibility score was calculated for the 5,168 participants with complete proteomic data at COPDGene’s visit 2 (considered the baseline visit for the present study).

FEV₁ Trajectory Groups

The FEV₁ trajectory groups used in this study were AD (with normal peak lung function) and normal decline (ND; with normal peak) (Figure 1). These were adapted from previously developed and published trajectories of FEV₁ percent predicted (7). In brief, longitudinal spirometry measurements (up to six time points over 30 yr) were used to generate trajectories of FEV₁ percent predicted using a group-based trajectory modeling approach (SAS PROC TRAJ). For the purposes of this study, participants with FEV₁ trajectories previously named “ideal,” “preserved good,” and “preserved impaired” lung health were combined into one trajectory called “normal decline” (normal peak with ND; n = 2,332). The previously named “worsening lung health” trajectory was renamed “accelerated decline” (normal peak with AD; n = 138). Participants with “persistently poor lung health” trajectory (n = 46) were excluded from analyses.

Quantification of the Human Proteome

For CARDIA participants, plasma samples drawn at the Year 25 visit were analyzed for protein identification using the SomaLogic SomaScan version 4.1 assay, using 7,335 SOMAmer aptamers to identify 6,609 unique human proteins. Proteomic identification was performed using the SomaScan version 4.0 assay in COPDGene, as previously described (12), and the Olink Explore 1536 assay in UKBB. The initial analyses of individual proteins associated with AD trajectory were performed in CARDIA using all 7,335 aptamers. However, the susceptibility score was derived using the 1,082 unique human proteins that overlapped (by Entrez Gene identifier) across proteomic assays to facilitate analyses of clinical outcomes in the two validation cohorts. For SOMAmer aptamers that targeted the same protein at different epitopes, we used the aptamer with the highest correlation with the matching Olink protein based on published data, when available (13).

Outcomes

Our outcomes of interest were incident COPD, respiratory exacerbations, respiratory death, and all-cause death. Incident COPD was only examined in UKBB, given the large proportion of participants with prevalent COPD in COPDGene at the time of proteomic measurement. Data on respiratory exacerbations were only available in COPDGene. In COPDGene, participant vital status was monitored through regular outreach to participants or their designated proxies as well as through searches of the national death indexes (14). In UKBB, death events were identified using death registry data, and respiratory death and incident COPD were defined by Phecode/International Classification of Diseases, 10th Revision, codes (Table E1) (15). In COPDGene, respiratory deaths were adjudicated by two reviewers evaluating and classifying deaths as respiratory, cardiovascular, cancer, or other (16). Exacerbations were defined in COPDGene as an episode of increased cough, phlegm, or shortness of breath that lasted >48 hours and required treatment with systemic steroids, antibiotics, or both (17). Severe exacerbations were defined as an exacerbation requiring an emergency room visit or hospitalization. In COPDGene, we also examined rate of FEV₁ decline, defined as the difference in post-bronchodilator FEV₁ percent predicted from visit 2 (time of proteomic measurement) to visit 3, divided by 5 years. Lung function measurements were not available in this study’s UKBB dataset.

Statistical Analysis

Proteomic correlates of AD trajectory

We constructed multivariable logistic regression models in CARDIA to measure associations between each of the 7,335 SOMAmer protein aptamers (log₂-transformed) and AD trajectory (vs. ND trajectory), adjusted for age, sex, center, and Year 20 predicted FEV₁ to identify proteins that are independent of important covariates. Adjustment for Year 20 FEV₁ was done to isolate the associations with lung function trajectory from the associations with cross-sectional lung function. Given the goal of discovering a protein signature that reflects respiratory susceptibility and encompasses the effects of traditional respiratory risk factors, rather than establishing a causal relationship between protein and FEV₁ decline, minimal covariate adjustment was done at this stage. We controlled for type I error using a false discovery rate (FDR) approach (Benjamini-Hochberg) with a 5% threshold. We performed gene ontology–based enrichment analysis for proteins passing 5% FDR using the DAVID (Database for Annotation, Visualization and Integrated Discovery) platform (18). We also examined ranked tissue-specific gene expression for significant proteins using Genotype-Tissue Expression (GTEx) database (19) and compared expression by each tissue type using the R package singscore (20).

Proteomic susceptibility score calculation

To derive our susceptibility score in CARDIA, we examined only the 1,082 proteins shared among the three cohorts. We performed a two-step reduction in the number of proteins included in the score: 1) logistic regression to identify protein correlates of AD trajectory that would all be independent of important clinical covariates age, sex, center, and Year 20 predicted FEV₁; and 2) proteins that were significantly associated with AD trajectory at an FDR of 5% were included thereafter in least absolute shrinkage and selection operator (LASSO) modeling to identify a parsimonious group of proteins that are associated with AD trajectory. LASSO is a common supervised machine learning technique that removes collinear predictors and promotes sparsity within the model. LASSO was performed with fivefold cross-validation to optimize model hyperparameters. Proteins were log₂-transformed and scaled to zero mean and unit variance, which is standard preprocessing for LASSO. The resulting LASSO model β-coefficients and intercept were used to generate a susceptibility score equation as follows: Susceptibility score = Intercept + (Coefficient₁ × Protein₁ level) + (Coefficient₂ × Protein₂ level) +…+ (Coefficient_X × Protein_X level). The susceptibility score was calculated for each participant in the three cohorts, and it was used as a continuous variable and categorized into quartiles for subsequent analyses. For proteins in the susceptibility score, gene expression within lung tissue was determined using the GTEx database (19).

To examine whether the susceptibility score was reflective of FEV₁ decline independent of tobacco smoke exposure, we stratified the CARDIA cohort by Year 25 smoking status and used a linear mixed effects model to test the association between susceptibility score and yearly change in FEV₁ percent predicted from Year 5 to Year 30, adjusted for age, sex, race, center, Year 25 body mass index (BMI), and Year 25 smoking pack-years. Year 5 was chosen rather than Year 0, given that in the CARDIA cohort, mean FEV₁ percent predicted generally peaks at Year 5 and declines linearly thereafter (Figure E2).

Clinical Outcomes

To test the association of the susceptibility score with clinical outcomes, we modeled time-to-event outcomes (all-cause mortality, respiratory mortality, incident COPD) using multivariable Cox proportional hazards regression models, rate of FEV₁ percent predicted decline using multivariable linear regression, odds of incident COPD exacerbation using multivariable logistic regression, and frequency of COPD exacerbations using multivariable zero-inflated negative binomial regression. Models were performed separately in each cohort. All models were complete-case analyses and adjusted for age, sex, self-identified race, BMI, smoking status, and pack-years at the time of proteomics. Given differences in available data across cohorts, we also adjusted for self-reported history of asthma and post-bronchodilator FEV₁ percent predicted (COPDGene) or self-reported respiratory disease (UKBB) and for income (COPDGene) or Townsend deprivation index (21) (UKBB). We visualized all-cause mortality across quartiles of the susceptibility score using Kaplan-Meier methods. In sensitivity analysis, we restricted the cohorts to only participants without lung disease at the time of proteomic measurement and examined the association between susceptibility score and all-cause mortality, respiratory mortality, and respiratory exacerbations. In COPDGene, this was defined by an FEV₁/FVC >0.70 and FEV₁ ≥80% predicted, and in UKBB, this was defined by self-reported or physician-diagnosed COPD, asthma, emphysema, chronic bronchitis, or respiratory failure. Additional methods are detailed in Appendix E1.

Results

Proteins Associated with an Accelerated Decline FEV₁ Trajectory

There were 138 participants with an AD trajectory and 2,332 participants with an ND trajectory in the CARDIA cohort (Figure 1). The characteristics of the CARDIA participants with these trajectories are shown in Table 1. In logistic regression of AD versus ND trajectory, 413 proteins were significant in univariate analysis, and 48 proteins (20 positively and 28 negatively associated with AD trajectory) were significant in multivariable analysis (Figure E3, Tables E2 and E3). Analysis of the tissue-specific gene expression activity of the 20 proteins positively associated with AD trajectory revealed that these proteins’ gene transcripts were most commonly expressed in adipose and lung tissue (Figure E4). Gene ontology analysis of the 48 proteins positively and negatively associated with AD trajectory demonstrated overrepresentation of proteins involved in extracellular components and in several different biological processes, including humoral immune response and defense response to bacterium (Figure E5).

Table 1.

CARDIA Cohort Characteristics, by FEV₁ Trajectory Group

Characteristic	Normal Decline Trajectory	Accelerated Decline Trajectory
No. of subjects	2,332	138
Age, yr, mean (SD)	50 (4)	49 (4)
Sex, n (%)
Male	1,017 (44%)	49 (36%)
Female	1,315 (56%)	89 (64%)
Race, n (%)
Black	1,034 (44%)	85 (62%)
White	1,298 (56%)	53 (38%)
Smoking status, n (%)
Never	1,463/2,299 (64%)	64/135 (47%)
Former	523/2,299 (23%)	21/135 (16%)
Current	313/2,299 (14%)	50/135 (37%)
Pack-years, mean (SD)	4.8 (9.5)	9.4 (12.2)
BMI, kg/m2, mean (SD)	29.9 (6.8)	34.9 (9.4)
Baseline FEV1 % predicted, mean (SD)	97.7 (12.6)	98.0 (13.0)
Baseline FEV1/FVC ratio, mean (SD)	83.2 (6.2)	82.2 (6.5)
Year 20 FEV1 % predicted, mean (SD)	94.1 (15.2)	80.5 (16.5)
Year 20 FEV1/FVC ratio, mean (SD)	0.79 (0.06)	0.75 (0.10)
Year 30 FEV1 % predicted, mean (SD)	93.3 (16.2)	66.8 (16.9)
Year 30 FEV1/FVC ratio, mean (SD)	0.77 (0.06)	0.70 (0.12)
Change in FEV1 % predicted*, mean (SD)	−4.3 (10.0)	−30.9 (12.3)
Change in FEV1 (mL/year)†, mean (SD)	−26.6 (11.2)	−50.4 (13.8)

Definition of abbreviation: BMI = body mass index.

^*Absolute change from Year 0 to Year 30.

^† Change per year from Year 0 to Year 30. Characteristics from Year 25 are displayed, unless otherwise specified.

Proteomic Risk Score of Increased Respiratory Susceptibility

Our susceptibility score was derived in the CARDIA cohort by LASSO regression and was composed of 32 proteins, with their LASSO coefficients listed in Table E4. Figure 2 displays the proteins in the susceptibility score together with their estimated lung-specific gene expression from the GTEx Portal, listing each protein’s function and/or associations with lung health and disease. The mean susceptibility score (calculated as a z-score) was 1.05 in the AD group and −0.06 in the ND group (difference = 1.11; 95% confidence interval [CI], 0.95–1.28), with distributions of susceptibility score depicted in Figure E6. In multivariable analyses stratified by smoking status at the time of proteomic measurement, we found that the susceptibility score was associated with 25-year decline in FEV₁ percent predicted, regardless of smoking status, including among never smokers, and with the largest association among current smokers (Table E5).

Figure 2.

Proteins in the susceptibility score and their function and/or associations with lung health and disease, ranked by lung-specific gene expression. Hierarchical clustering was generated using the Genotype-Tissue Expression portal based on lung-specific gene expression measured in median transcripts per million. Least absolute shrinkage and selection operator (LASSO) directionality refers to the direction of their association with accelerated decline lung function trajectory in the LASSO model. Protein function data were obtained from the UniProt database unless indicated by reference numbers (47–74).

Association of Susceptibility Score with Future Lung Function Decline

The UKBB and COPDGene cohort characteristics by quartile of susceptibility score are presented in Table 2. We first examined whether the susceptibility score was associated with future lung function decline over 5 years from the time of proteomic measurement in COPDGene. We found that participants with the highest quartile of susceptibility score had an absolute decline in post-bronchodilator FEV₁ percent predicted that was 1.8% (95% CI, 0.09–3.5%) greater than that in the lowest quartile (Table E6). There was no statistically significant difference in FEV₁ percent predicted decline between quartiles 1 and 2 or 3.

Table 2.

Characteristics of UK Biobank and COPDGene Cohorts, by Quartile of Proteomic Respiratory Susceptibility Score

Characteristic	U.K. Biobank				COPDGene
	Quartile 1, n = 8,750	Quartile 2, n = 8,750	Quartile 3, n = 8,750	Quartile 4, n = 8,750	Quartile 1, n = 1,298	Quartile 2, n = 1,299	Quartile 3, n = 1,291	Quartile 4, n = 1,280
Susceptibility score*	−1.40 (0.47)	−0.44 (0.20)	0.29 (0.23)	1.54 (0.72)	−1.29 (0.46)	−0.39 (0.18)	0.28 (0.21)	1.39 (0.70)
Age, yr*	53 (8)	56 (8)	58 (8)	60 (7)	64 (8)	66 (9)	66 (9)	66 (9)
Female, n (%)	4,891 (56)	4,719 (54)	4,618 (53)	4,639 (53)	655 (50)	625 (48)	643 (50)	635 (50)
Race, n (%)
Asian	198 (2.3)	183 (2.1)	176 (2.0)	173 (2.0)	0	0	0	0
Black	324 (3.7)	159 (1.8)	155 (1.8)	114 (1.3)	342 (26)	336 (26)	377 (29)	448 (35)
Mixed	89 (1.0)	54 (0.6)	47 (0.5)	48 (0.5)	0	0	0	0
Unknown/other	186 (2.1)	148 (1.7)	121 (1.4)	114 (1.3)	0	0	0	0
White	7,953 (91)	8,206 (94)	8,251 (94)	8,301 (95)	956 (74)	963 (74)	914 (71)	832 (65)
BMI*	24.9 (3.3)	26.6 (3.7)	28.1 (4.3)	30.3 (5.6)	26.4 (4.7)	28.4 (5.4)	29.8 (6.3)	31.5 (7.6)
Smoking status, n (%)
Current	400 (4.6)	606 (6.9)	924 (11)	1,731 (20)	338 (26)	469 (36)	542 (42)	634 (50)
Former	2,590 (30)	2,984 (34)	3,302 (38)	3,308 (38)	960 (74)	830 (64)	749 (58)	646 (50)
Never	5,754 (66)	5,148 (59)	4,513 (52)	3,700 (42)	—	—	—	—
Missing	6 (<1)	12 (<1)	11 (<1)	11 (<1)	—	—	—	—
Pack-years*	3 (8)	5 (12)	8 (15)	13 (21)	37 (21)	44 (22)	46 (24)	49 (26)
Deprivation index*†	−1.5 (3.1)	−1.4 (3.1)	−1.2 (3.1)	−0.7 (3.3)	—	—	—	—
Income, n (%)
<$15,000	—	—	—	—	253 (19)	310 (24)	374 (29)	466 (36)
$15,000–$35,000	—	—	—	—	278 (21)	298 (23)	315 (24)	297 (23)
$35,000–$50,000	—	—	—	—	186 (14)	194 (15)	167 (13)	142 (11)
$50,000–$75,000	—	—	—	—	206 (16)	177 (14)	147 (11)	108 (8.4)
>$75,000	—	—	—	—	213 (16)	156 (12)	118 (9.1)	77 (6.0)
Decline to answer	—	—	—	—	162 (12)	164 (13)	170 (13)	190 (15)
Respiratory disease‡, n (%)	1,041 (12)	1,129 (13)	1,194 (14)	1,536 (18)	—	—	—	—
FEV1/FVC ratio <0.7, n (%)	—	—	—	—	447 (34)	579 (45)	651 (50)	639 (50)
FEV1 % predicted*	—	—	—	—	86 (24)	79 (25)	74 (24)	72 (23)

Definition of abbreviation: BMI = body mass index.

^*Data are presented as mean (SD).

^† Townsend deprivation index is a census-based index of area material deprivation, with a more positive score indicating more material deprivation.

^‡ Respiratory disease includes self-report of asthma, emphysema, chronic obstructive pulmonary disease, chronic bronchitis or respiratory failure, and physician-diagnosed chronic obstructive pulmonary disease, asthma, emphysema, or chronic bronchitis at baseline visit.

Association of Susceptibility Score with All-Cause Mortality and Respiratory Mortality

In multivariable models, a 1-SD higher susceptibility score was significantly associated with all-cause mortality in both cohorts: 56% higher risk in UKBB (hazard ratio [HR], 1.56; 95% CI, 1.50–1.61) and 75% higher in COPDGene (HR, 1.75; 95% CI, 1.63–1.88). Each 1-SD higher susceptibility score was significantly associated with over two times the risk of respiratory death in UKBB (HR, 2.39; 95% CI, 2.16–2.64) and 81% higher risk in COPDGene (HR, 1.81; 95% CI, 1.32–2.47). We also found a dose–response relationship between susceptibility score quartile and risk of all-cause death and respiratory death (Figure 3). Unadjusted cumulative event plots for all-cause mortality in both cohorts are displayed in Figure 4. The number of participants, deaths, and follow-up time for each cohort are shown in Table E1. Causes of death for the two cohorts are shown in Tables E7 and E8.

Figure 3.

Mortality, incident COPD, and incident respiratory exacerbation by quartile of proteomic respiratory susceptibility score. (A) Adjusted hazard ratio (HR) for all-cause death and respiratory death ascertained in U.K. Biobank (UKBB) and COPDGene. (B) Adjusted HR for incident chronic obstructive pulmonary disease (COPD) in UKBB. (C) Adjusted odds ratio for one or more exacerbations and one or more severe exacerbations. All analyses were adjusted for age, sex, self-identified race, body mass index (BMI), smoking status, and pack-years at the time of proteomic measurement. In COPDGene, additional adjustment was made for self-reported asthma, post-bronchodilator FEV₁ percent predicted, and income. UKBB analyses were adjusted for self-report of current respiratory disease (COPD, emphysema, chronic bronchitis, asthma, or respiratory failure) and Townsend deprivation index. The Townsend deprivation index estimates deprivation within a census area and comprises four variables: unemployment, non–car ownership, non–home ownership, and household overcrowding. Analyses in COPDGene were additionally adjusted for platelet count and white blood cell count on the basis of prior internal quality control. CI = confidence interval.

Figure 4.

Cumulative incidence plots of all-cause mortality by quartile of proteomic respiratory susceptibility score. Unadjusted cumulative incidence of all-cause death by quartile of proteomic susceptibility score in the U.K. Biobank (top) and COPDGene (bottom) cohorts. Shaded areas around each line represent the 95% confidence interval.

Association of Susceptibility Score with Respiratory Disease and Exacerbations

We used the unique set of large-scale population data in UKBB to understand the association between the susceptibility score and incident respiratory disease. Each 1-SD higher susceptibility score was associated with increased risk of incident COPD (HR, 1.84; 95% CI, 1.71–1.98). We used detailed data on respiratory exacerbations in the smoking-enriched population in COPDGene to examine associations between the susceptibility score and respiratory morbidity. In COPDGene, each 1-SD higher susceptibility score was associated with higher odds of one of more future respiratory exacerbations (odds ratio [OR], 1.10; 95% CI, 1.02–1.19) and future severe exacerbations (OR, 1.17; 95% CI, 1.08–1.27). Furthermore, the susceptibility score was associated with higher frequency of future exacerbations (incidence rate ratio, 1.28; 95% CI, 1.19–1.38) and severe exacerbations (incidence rate ratio, 1.33; 95% CI, 1.20–1.48). Associations by susceptibility score quartile are displayed in Figure 3.

Susceptibility Score among Those without Lung Disease at Baseline

In multivariable sensitivity analyses, we restricted the cohorts to those without lung disease at baseline and found results similar to those with the entire cohort. There was a 52% higher risk of all-cause death in UKBB (HR, 1.52; 95% CI, 1.46–1.58) and an 89% higher risk in COPDGene (HR, 1.89; 95% CI, 1.61–2.22) associated with 1-SD higher susceptibility score. In UKBB, this was also associated with 2.39 times the risk of respiratory death (HR, 2.34; 95% CI, 2.07–2.66). Among those without baseline respiratory disease in COPDGene, there was only one respiratory death, and therefore an HR could not be calculated. We also found in COPDGene that 1-SD higher susceptibility score was associated with higher odds of one of more future respiratory exacerbations (OR, 1.13; 95% CI, 1.00–1.28) but not risk of future severe exacerbations (OR, 1.10; 95% CI, 0.94–1.28). Associations by susceptibility score quartile are displayed in Table E9.

Discussion

We used data from three diverse cohorts with varied smoking histories to characterize a proteomic risk score of increased respiratory susceptibility that was associated with future respiratory morbidity and mortality. A susceptibility score in the highest quartile was associated with a 54% higher odds of respiratory exacerbation requiring hospitalization, 4 times higher odds of incident COPD, and 3.8 to 6 times higher odds of respiratory mortality than the lowest quartile. Many of the proteins included in the score are highly expressed in lung tissue. Some proteins have well-established roles in lung health and immunity, whereas others are novel and may hold new insights into respiratory health. This tool allows the identification of individuals with increased susceptibility to future respiratory morbidity and mortality without requiring years of clinical observation.

Precision medicine omic studies have made important advances in predicting who may be at highest risk of developing chronic lung disease. Several genome-wide association studies have discovered polygenic risk scores associated with risk of lung function decline and future COPD (22–24). Proteomics is a complementary approach that integrates genetic determinants of lung health with host response to exposures. Because proteomics represents downstream perturbations caused by not only genetics but also the environment, proteomic biomarkers of susceptibility and impaired respiratory health may represent potentially modifiable risk factors and offer opportunities for disease intervention. As expected, we found that our susceptibility score was associated with traditional risk factors for impaired respiratory health and chronic lung disease, including smoking, asthma, and increased BMI, as well as social determinants of health, such as measures of lower socioeconomic status and race (exposure to racism or racial health inequities).

Notably, we found that the susceptibility score was associated with FEV₁ decline in CARDIA, regardless of smoking status, including among never smokers, suggesting it may identify disease risk outside of tobacco alone. The susceptibility score was also associated with incident COPD, respiratory mortality, and all-cause mortality independent of smoking, age, sex, race, asthma diagnosis, BMI, and socioeconomic status. Furthermore, we demonstrated that a proteomic risk score derived in a cohort at mean age 50 can predict respiratory outcomes in individuals aged 45 to 75 years. This suggests that our score is generalizable to an age range common for COPD diagnosis as well as to earlier ages when disease interception may be more feasible (25).

Although the application of proteomic approaches to understand decline in lung function is not entirely new, prior studies have been limited to short-interval (1–7-yr) FEV₁ declines and without clear implications for life course trajectories, especially at an early modifiable period of impaired health (26, 27). CARDIA, a longitudinal study with over 35 years of follow-up and validated lung function trajectories (7), offers unique advantages. Among 2,470 CARDIA participants included in initial trajectory analyses, >97% had four to six FEV₁ measurements over 30-year follow-up, beginning at ages 18–30 years. These longitudinal measurements facilitate precise estimates of peak lung function and long-term FEV₁ change to generate a robust life course trajectory of FEV₁ decline on which to define respiratory susceptibility. Prior studies have shown that accelerated FEV₁ decline trajectory is a phenotype associated with increased risk of respiratory and all-cause death (5, 6). Our study fills an important knowledge gap by identifying a blood biomarker surrogate of respiratory susceptibility based on people studied for decades in CARDIA and confirmed to have AD. Thus, although the proteomic score is independently associated with short-term FEV₁ decline in COPDGene, its strength is instead in identifying respiratory susceptibility even in those without decades of longitudinal spirometry.

Prior studies have examined individual biomarkers, or combinations of selected biomarkers, and demonstrated their association with accelerated FEV₁ decline, COPD hospitalizations, and death (28–31). However, broad, discovery-based proteomic sampling has the advantage of finding novel predictors understudied in the context of lung disease. The proteins uncovered in this study not only confirmed well-established pathways related to COPD development but also new associations with lung function decline and respiratory mortality not widely recognized. As depicted in Figure 2, mechanisms of fibrosis, inflammation, and immunity were prominently implicated by proteins in the susceptibility score, including membrane-associated PLA2 (secreted into alveolar space by alveolar macrophages and central to inflammatory responses to lung injury [32–34]), epidermal growth factor receptor (key regulator of airway mucus production and secretion and implicated in the pathogenesis of asthma, COPD, and lung cancer [35–37]), WAP four-disulfide core domain protein 2 (mediator of myofibroblast activity that is associated with idiopathic pulmonary fibrosis and other forms of progressive pulmonary fibrosis), CDCP1 (negative regulator of transforming growth factor-β1 signaling, a pathway implicated in airway fibrosis and hyperreactivity and myofibroblast differentiation [38, 39]), Dickkopf-related protein 3 (regulates Wnt/β-catenin pathways and is involved in inflammation in the lung [40, 41]), cystatin-C (cysteine protease inhibitor associated with lung function, COPD, pulmonary hypertension, and acute respiratory distress syndrome mortality [42–45]), and LILRB4 (expressed on macrophages and upregulated in human COPD lung tissue and in a mouse model of emphysema [46]). Given the pleotropic effects of inflammation and fibrosis pathways in multiorgan function, we cannot definitively assert the pulmonary specificity of our susceptibility score. However, the expression of key proteins at a transcriptional level in lung tissue coupled with the strong association with pulmonary outcomes across studies reinforces clinicopathologic validity. In deriving the susceptibility score, we limited the analysis to proteins that overlapped between the SomaScan and Olink proteomic assays. Although this may have excluded some potentially biologically relevant proteins, the resultant susceptibility score remains a significant and more generalizable predictor of respiratory disease and morbidity.

Limitations

This study represents a unique union of human proteomic data with longitudinal spirometry and long-term clinical outcomes among those with and without overt lung disease. However, several limitations warrant mention. Participants in CARDIA only had one spirometry measurement after the collection of blood for proteomic measurements, limiting any assertion regarding the ability of the proteomic scores to predict incident decline in FEV₁. In addition, although we demonstrated in COPDGene that a high susceptibility score is associated with faster short-term FEV₁ decline than a low susceptibility score, there was no comparable validation cohort with life course measures of lung function and similar proteomic assay to confirm the susceptibility signature and score as a predictor of AD trajectory. However, predicting trajectory was not the goal of our study. Instead, we aimed to use the AD in FEV₁ trajectory as a measure of respiratory susceptibility, and we captured the proteomic perturbations associated with respiratory susceptibility as a predictor of future adverse pulmonary outcomes in multiple cohorts. Longitudinal cohorts of healthy adults focused on the development of lung disease are needed to evaluate whether the susceptibility score can also predict AD trajectory. In addition, inherent limitations of observational studies including unmeasured confounding and competing risk from all-cause mortality may have impacted the estimates of risk of other outcomes, such as respiratory death, incident lung disease, and exacerbations. Another limitation is that both the SomaScan and Olink proteomic assays are semiquantitative, and the specificity of any one aptamer or antibody for the target protein in the derivation sample is likely variable. Nonetheless, even while using two different proteomic platforms, the susceptibility score consistently remained associated with respiratory outcomes.

Future studies with absolute quantification of the proteins in the susceptibility score are required for translation to clinical practice. However, the potential applications of a blood test for early detection of lung disease risk, before irreversible lung damage and marked lung function decline have occurred, could have enormous impact. Furthermore, a panel of protein biomarkers such as the susceptibility score holds the potential to capture not only genetic risk but also potentially modifiable risk from environmental and behavioral factors. Still, whether the proteins in the susceptibility score represent intervenable pathways of lung-specific pathology remains an open question. Mechanistic studies to better understand the roles of these proteins in lung tissue are warranted to address this important gap. This translation from observations and proteomic biomarker discovery to inferring a causal role in disease pathobiology represents a critical step to identify potentially druggable targets for the future interception of impaired respiratory health.

Conclusions

AD in FEV₁ trajectory has been recognized as a morbid clinical phenotype highly susceptible to chronic respiratory disease; however, it has been unclear so far how to apply this knowledge to clinical practice, given the impracticalities of performing repeated spirometry measurements in healthy people. The present study identifies a human proteomic signature of respiratory susceptibility and demonstrates its ability to identify individuals at risk for future COPD, hospitalization for respiratory exacerbation, and death caused by respiratory disease. The proteins identified in this study show promise as biomarkers for impaired respiratory health. Further study to elucidate their utility as modifiable targets for the prevention and interception of chronic lung disease in at-risk populations is warranted.