Proteomic biomarkers of emphysema-predominant and non-emphysema-predominant chronic obstructive pulmonary disease

Summary

Background

Chronic Obstructive Pulmonary Disease (COPD) is a complex and heterogeneous disease. Emphysema-predominant and non-emphysema predominant COPD are two major disease subtypes capturing important aspects of COPD heterogeneity. Molecular differences between these COPD subtypes are unknown.

Methods

We assessed plasma proteomic associations (using SomaScan) with emphysema-predominant vs. non-emphysema predominant COPD subtypes in COPDGene; replication of significant associations was performed in SPIROMICS. We performed pathway analyses on COPD subtype plasma proteomic associations and used weighted gene correlation network analysis to find COPD subtype-associated protein correlation networks. We tested previously reported COPD genetic variants for association with COPD subtypes and COPD subtype-associated proteomic biomarkers.

Findings

One hundred and twenty-four proteins were significantly associated with COPD subtypes in COPDGene, with 64 proteins (65 SOMAmers) validated in SPIROMICS. Higher correlations were observed between proteomic biomarkers with greater expression levels in non-emphysema predominant participants with COPD. Cell adhesion, collagen-containing extracellular matrix, and epithelial mesenchymal transition were biological pathways enriched for COPD subtype proteomic associations. One COPD subtype-associated correlation network module was identified, including highly connected proteomic biomarkers like PXDN and EFNA2. We observed significant genetic effects on COPD subtypes for rs2579762 in LRMDA and on COPD subtype-associated proteomic biomarkers including sRAGE and Ganglioside GM2 Activator.

Interpretation

We identified and replicated multiple plasma proteomic biomarkers associated with emphysema-predominant vs. non-emphysema predominant COPD. Pathway analyses, correlation-based network analyses, and genetic association analyses of these proteins may provide insight into the molecular heterogeneity of COPD.

Funding

National Heart, Lung, and Blood Institute (NIH).

Research in Context.

Evidence before this study

Chronic obstructive pulmonary disease (COPD) is a complex and heterogeneous condition characterised by persistent airflow limitation. COPD subtyping could lead to improvements in disease diagnosis, progression prognostication, and treatment strategy optimisation. Multiple previous approaches to examine COPD heterogeneity have converged upon a subtyping distinction based on quantitative emphysema assessed by chest CT scan. Subjects with significant airflow obstruction (FEV1 < 80% predicted with FEV1/FVC < 0.7) and greater than or equal to 10% computed tomography (CT) quantitative densitometric emphysema (percentage of low attenuation areas less than −950 Hounsfield units (LAA-950)) have been classified as emphysema predominant COPD (EPD), while non-emphysema-predominant disease (NEPD) has been defined by significant airflow obstruction with <5% LAA-950. This EPD/NEPD COPD subtyping classification is associated with multiple previously applied COPD subtype definitions and reflects COPD progression patterns.

Previous proteomic studies have identified plasma biomarkers associated with COPD affection status, lung function decline, and emphysema.

Added value of this study

We used a widely reported aptamer-based proteomics assay (SOMAScan) to identify molecular differences between emphysema and non-emphysema predominant COPD in plasma. We recognised 124 plasma protein biomarkers associated with EPD vs. NEPD in the COPDGene study, and we were able to replicate most of these associations in the SPIROMICS study. These proteomic biomarkers are enriched in biological functions including cell adhesion, collagen-containing extracellular matrix, and epithelial mesenchymal transition. Correlation-based network analyses identified highly connected proteomic biomarkers within a COPD subtype-associated correlation module.

Implications of all the available evidence

Multiple plasma protein biomarkers distinguish EPD and NEPD subtypes of COPD, suggesting that these COPD subtypes have different pathogenetic mechanisms. This information could eventually lead to more accurate disease diagnosis, disease progression prognostication, and treatment optimisation.

Introduction

Chronic Obstructive Pulmonary Disease (COPD) is a complex and heterogeneous disease defined by persistent airflow obstruction.^1,2 The identification of subtypes within the COPD syndrome is important for more accurate disease diagnosis, disease progression prognostication, and treatment optimisation. However, current diagnostic approaches do not account for heterogeneity in COPD.³

Multiple approaches have been used to examine COPD heterogeneity using clinical, physiological, and imaging phenotypes. Castaldi et al. identified four clusters associated with clinical characteristics, including exacerbations, and COPD-associated genetic variants using K-means cluster analysis.⁴ Kinney et al. recognised five factors/disease axes to represent COPD pathophysiologic heterogeneity by factor analysis integrating pulmonary function testing and computed tomography imaging results.⁵ Ross et al. identified four lung function trajectories with significant genetic associations using a data-driven approach.⁶ Protein biomarkers have also been used in cluster analysis to identify COPD subtypes.⁷ However, there remains little consensus about COPD subtyping, and transferability of unsupervised clustering solutions across different study populations is limited.⁸

Emphysema is an important clinical parameter capturing COPD heterogeneity. Patients with more emphysema have, on average, lower lung function, longer duration of disease, higher mortality,^9,10 and lower exercise capacity.¹¹ Castaldi et al. assessed the heterogeneity of COPD through emphysema severity¹²; emphysema-predominant and non-emphysema predominant COPD were classified using spirometric and quantitative imaging criteria. In participants with COPD, emphysema-predominant disease (EPD) was defined as significant airflow obstruction (FEV1 < 80% predicted with FEV1/FVC < 0.7) with ≥10% computed tomography (CT) quantitative densitometric emphysema (percentage of low attenuation areas less than −950 Hounsfield units (LAA-950)), while non-emphysema-predominant disease (NEPD) was defined by significant airflow obstruction with <5% LAA-950. This EPD/NEPD COPD subtyping classification was strongly associated with several previously applied COPD subtype definitions and reflected different COPD progression patterns.¹² However, the molecular differences between emphysema-predominant and non-emphysema predominant COPD have not been assessed.

Proteomic analysis can reveal the interactions, functions, and structures of proteins¹³ and provide robust molecular profiles for a systematic understanding of disease pathogenesis.¹⁴ Plasma proteomics reflects the molecular status in the peripheral circulation for systemic disease assessments, monitoring, and screening.¹⁵ Several circulating proteomic biomarkers have been previously associated with different aspects of COPD: 1) COPD affection status biomarkers including C-reactive protein^16,17; 2) FEV1 biomarkers including interleukin 6¹⁸; 3) FEV1/FVC ratio biomarkers including lipocalin-1¹⁹; and 4) emphysema biomarkers including sRAGE.^20,21

Several large-scale commercial proteomic panels have been developed; we used SomaScan, an aptamer-based proteomics assay capturing thousands of human proteins in various biological materials with high sensitivity and specificity.²² We hypothesised that plasma proteomics analysis would reveal molecular differences between emphysema-predominant and non-emphysema-predominant COPD.

Methods

Study population

The COPDGene Study is a multi-centre observational study designed to investigate the pathobiological mechanisms, heterogeneity, and progression of Chronic Obstructive Pulmonary Disease (COPD).²³ The heterogeneity of COPD has been observed at different levels including physiology, imaging, clinical course, and response to therapeutic intervention.²⁴ The COPDGene Study has enrolled more than 10,000 smokers with or without COPD across different COPD GOLD spirometry grades. Clinical phenotypes including chest computed tomography (CT) phenotypes have been characterised to evaluate emphysema, gas trapping, and airway wall thickening.²³ The first participants in COPDGene were enrolled in 2007 (Phase 1), and baseline enrolment ended in 2011. Non-Hispanic White and Black subjects with at least 10 pack-years of smoking and self-reported age of 45–80 at enrolment were eligible to participate. A set of non-smoking controls was also enrolled, but they were not included in this project. Five-year follow-up data collection was obtained (Phase 2 cohort, 2013–2017) which included fresh frozen plasma collection. Subjects who participated in Phase 2 (2013–2017) of the study were selected for plasma proteomic analysis. Participants with emphysema-predominant COPD, non-emphysema predominant COPD, or controls from the COPDGene Phase 2 cohort were selected as follows:

• I.Control population: FEV1 ≥ 80% predicted, FEV1/FVC ≥ 0.7 based on post-bronchodilator spirometry (pre-bronchodilator spirometry was used if post-bronchodilator spirometry was not available), pack-years ≥ 10 and LAA-950 < 5%
• II.Emphysema-predominant COPD: FEV1 < 80% predicted, FEV1/FVC < 0.7 based on post-bronchodilator spirometry (pre-bronchodilator spirometry was used if post-bronchodilator spirometry was not available), pack-years ≥ 10 and LAA-950 ≥ 10%
• III.Non-emphysema predominant COPD: FEV1 < 80% predicted, FEV1/FVC < 0.7 based on post-bronchodilator spirometry (pre-bronchodilator spirometry was used if post-bronchodilator spirometry was not available), pack-years ≥ 10 and LAA-950 < 5%

SPIROMICS (SubPopulations and InteRmediate Outcome Measures In COPD Study) supports the prospective collection and analysis of phenotypic, biomarker, genetic, genomic, and clinical data from subjects with COPD to identify subpopulations and intermediate outcome measures. We included subjects with a minimum of 20 pack-years of smoking from SPIROMICS using the same definitions of emphysema-predominant COPD, non-emphysema predominant COPD, and control subjects as in COPDGene to assess replication of proteomic associations.

The proteomics analysis workflow is shown in Fig. 1.

Fig. 1.

CONSORT diagram for COPD subtype proteomics analysis.

Ethics

All study participants provided written informed consent, and studies were approved by local Institutional Review Boards. The current study was approved by the Mass General Brigham institutional review board (IRB #2007P000554).

Omics data preprocessing

Plasma samples were sent to SomaLogic for protein quantification using SomaLogic SomaScan® version 4.0 (5.0K) assay for human plasma in the COPDGene cohort and version 4.1 (7.0K) assay in SPIROMICS. For the COPDGene cohort, initially there were 5285 SOMAmers, 4979 of which assayed human proteins representing 4860 proteins or protein subunits (4727 UniProt IDs). SomaLogic standardised the SomaScan® data per their protocol using the approach in Serban et al.²⁵ To control variability between sample batches and individual assays, median signal normalisation was performed using adaptive normalisation by maximum likelihood; plate scaling and calibration of SOMAmer analytes were also applied. Before data analysis, logarithmic transformation with base 2 was performed for normalisation. For the SPIROMICS cohort, there were 7596 SOMAmers representing 6626 human proteins or protein subunits (6401 UniProt IDs). Among all 4979 SOMAmers from SomaLogic SomaScan® version 4.0 (5.0K) assays detecting human proteins, there were 46 SOMAmers detecting protein complexes that consist of multiple proteins. As for 7596 SOMAmers from version 4.1 (7.0K) assays, there were 82 SOMAmers detecting multiple proteins in protein complexes. In 125 selected COPD subtype-associated SOMAmers, one SOMAmer (3359_11, targeting Cyclin-dependent kinase 8: Cyclin-C complex) detects a protein complex.

For downstream analyses, proteomic biomarker identification was performed at the SOMAmer level and annotated with protein name and UniProt ID. Selected proteomic biomarker numbers were summarised at both SOMAmer and UniProt ID levels. Proteomic associations for downstream pathway analyses including gene set enrichment analyses and gene ontology enrichment analyses were summarised at the unique HGNC (Human Gene Organisation Gene Nomenclature Committee) symbol level to avoid redundancy.

Genotyping data in the COPDGene cohort were obtained using the Human Omni Express chip (Illumina, San Diego, CA). Genotype imputation on the COPDGene cohorts was performed across the genome for the COPDGene cohort with the Haplotype Reference Consortium (HRC)²⁶ for non-Hispanic whites and 1000 Genomes Phase 1 v3 for Blacks, respectively. Eighty-two genome-wide significant COPD GWAS loci were previously reported in a genome-wide association analysis with 35,735 COPD cases and 222,076 controls from UK Biobank and the International COPD Genetics Consortium.²⁷ We also included rs28929474 (Z allele)²⁸ in SERPINA1 which was genotyped by Illumina HumanExome arrays when available or imputed from GWAS data if not available. We included genotyped or imputed data for 82 COPD GWAS loci and rs28929474 from control, emphysema-predominant COPD, and non-emphysema predominant COPD participants from the Phase 2 COPDGene cohort for further analysis.

Proteomic biomarker identification

We utilised linear regression models to identify proteomic biomarkers comparing emphysema-predominant vs. non-emphysema predominant COPD. Age, sex, race, pack-years, BMI, smoking status (current or former smokers), clinical centre, and cellular components (white blood cell counts and platelet counts) were introduced as covariates for linear regression models, as shown below:

$$ProteomicSOMAmers\sim Emphysema/Non-emphysemapredominantCOPD+age+sex+race+pack-years+BMI+currentsmokingstatus+clinicalcenter+cellularcomponents\left(platelets,whitebloodcells\right)$$

To evaluate whether our COPDGene cohort is large enough to detect significant proteomic associations, we performed power calculations using R package pwr v. 1.3-0 for the linear regression model. Minimal sample size was estimated with R² for linear regression model controlled at 0.3 and statistical power controlled at 0.9 to evaluate weak association detection capacity (variables with low explanatory levels). SPIROMICS proteomics data were generated from SomaScan 4.1 (7.0K) assay as replication. Proteomic biomarkers for each comparison were selected with false discovery rate (FDR) controlled at 0.05 in the COPDGene and SPIROMICS cohorts separately.

For additional comparisons between linear regression models with or without adjustment for FEV1% predicted as a sensitivity test, we also test their significant concordance in effect direction.

Pair-wise protein-protein Pearson's correlations for emphysema-predominant vs. non-emphysema predominant COPD proteomic biomarkers with higher expression levels in emphysema-predominant or lower expression levels in emphysema-predominant were compared.

Considering the epidemiological associations between COPD and cardiovascular diseases, we searched for overlap between COPD subtype biomarkers and previously reported cardiac biomarkers in two datasets: Gene2Phenotype Cardiac Gene Panel²⁹ and Cardiovascular Disease Atlas.³⁰

Pathway analysis

Gene set enrichment and gene ontology enrichment analyses were conducted to explore pathway-level differences in the pathogenesis of COPD subtypes. We performed gene set enrichment analysis using R package GAGE v.2.48.0³¹ with HALLMARK,³² KEGG,³³ and REACTOME³⁴ gene sets from Molecular Signatures DataBase (MSigDB).^35,36 Proteins were initially ranked based on the P-values (statistical significance) from linear regression models for emphysema-predominant vs. non-emphysema predominant COPD biomarker identification. We then integrated the negative log 10 transformed P-values with the direction of beta coefficients (negative for higher expression level in non-emphysema predominant COPD, and positive for higher expression level in emphysema-predominant COPD) to capture regulatory directions. Higher absolute values represent the most differentially detected proteins. Positive directions indicate proteins with higher expression level in emphysema-predominant COPD, while negative directions indicate proteins with higher expression level in non-emphysema predominant COPD.

We also performed gene ontology enrichment analysis using R package topGO v.2.50³⁷ with significant emphysema-predominant vs. non-emphysema predominant COPD proteomic biomarkers with FDR < 0.05 as genes of interest and all detected proteins from SomaScan assay as gene background.

Weighted gene correlation network analysis

We calculated proteomic residuals to remove effects of covariates including age, sex, race, pack-years, BMI, clinical centre, current smoking status, and cellular components (platelets and white blood cells). Weighted gene correlation network analysis (WGCNA) was performed on all available proteins (4979 SOMAmers, 4727 proteins at UniProt ID level) in COPDGene COPD subjects (emphysema-predominant and non-emphysema predominant COPD) using the R package WGCNA v1.70-3³⁸ with proteomic residuals after covariate adjustment. We set minimal module size at ten and optimised soft threshold (power) according to scale-free topology model fit and mean connectivity to evaluate connectivity between network nodes.

We estimated eigengene expression level for each identified module. The module eigengene is the first principal component of proteins from that module. Associations between module eigengenes and emphysema-predominant vs. non-emphysema predominant COPD were evaluated using linear regression models:

$$Moduleeigengene\sim Emphysema-predominantvs.Non-emphysemapredominantCOPD$$

Module membership is defined as the Pearson's correlation between module eigengene expression level and each single gene in the module, reflecting the module representativity and connectivity level of its gene components. Gene significance is defined as the association between each gene from the module and a specific trait. We evaluated the module representativity of components in the selected module by calculating Pearson's correlation coefficients between component expression levels and eigengene expression level.

Gene ontology enrichment analysis was performed on genes from COPD subtype associated modules using R package topGO v.2.50.³⁷

Phenotypic and proteomic genetic associations

In COPDGene, we summarised candidate SNPs including 82 COPD GWAS loci from Sakornsakolpat et al.²⁷ and rs28929474 (Z allele) in SERPINA1. To evaluate whether the COPDGene cohort is large enough to detect significant quantitative trait loci associations, we performed power calculations using R package pwr.

We tested COPD-associated SNPs with pairwise comparisons among 1) control population, 2) emphysema-predominant COPD, and 3) non-emphysema predominant COPD using additively coded genotypes. Logistic regression models were established comparing COPD subtypes and control participants with the following covariates: age, sex, pack-years, BMI, current smoking status, clinical centre, and top ten principal components of genetic ancestry, as shown below. Models were stratified by race (non-Hispanic White and Black participants), and results were combined by fixed effect meta-analyses using METAL.³⁹

$$Diseasestatus\sim COPDGWASSNPs+age+sex+pack-years+BMI+currentsmokingstatus+clinicalcentre+populationstratificationPCs$$

We performed quantitative trait locus analysis using MatrixEQTL⁴⁰ to reveal the associations between COPD genetic loci and proteomic biomarkers with age, sex, pack-years, BMI, current smoking status, clinical centres, top ten principal components of genetic ancestry, and top two proteomic principal components as covariates, as shown below:

$$COPDsubtypebiomarkers\sim COPDGWASSNPs+age+sex+pack-years+BMI+currentsmokingstatus+clinicalcentre+cellularcomponents\left(platelets,whitebloodcells\right)+populationstratificationPCs+proteinexpressionPCs$$

Genomic locations for candidate proteins and COPD GWAS loci were summarised based on GRCh37 human genome assemblies. We also combined QTL associations across ancestry groups using METAL.³⁹

Machine learning model for COPD subtype classification

To predict emphysema and non-emphysema predominant COPD subtypes, we created a random forest model. We utilised 10-fold cross-validation to tune parameters: number of trees and mtry with optimised prediction accuracy. We optimised a random forest model in the COPDGene proteomics cohort using 10 fold cross-validation evaluated by prediction accuracy. A feature selection approach was performed for each training process using the mutual information between each protein and COPD subtypes. The final model was established in the entire COPDGene proteomics cohort with the optimised number of trees and mtry. The final optimised random forest model was tested in SPIROMICS, and the feature importance from the final optimised model was also evaluated.

Statistics

Clinical characteristics

Quantitative variables were compared among control, emphysema-predominant COPD, and non-emphysema predominant COPD participants using independent Student's t-tests, while categorical variables were analysed using Chi-square tests. Chi-square tests were used since the expected number of observations was at least 5 for each cell of the contingency table analysis.

Sample size determination

With linear regression R² controlled at 0.3 (variables with low explanatory levels), we utilised R package pwr to estimate minimal sample size required for COPD subtype proteomic biomarker identification and quantitative trait loci recognition.

Proteomic biomarker identification

For linear regression models, we selected covariates based on known effects on SomaScan assay results: clinical centre and cellular components and similar covariates as have been used in other COPD clinical epidemiological studies: age, sex, pack-years, current smoking, and BMI. We used one-way ANOVA followed by the Student-Newman-Keuls (SNK) post-hoc test to identify differentially expressed proteins among EPD, NEPD, and control groups. This approach offers greater statistical power to detect group-specific differences compared to more conservative multiple comparison methods.

Sensitivity analyses

Sensitivity analysis was performed to assess the stability of the results by adjusting for FEV1% predicted. This helps determine whether the primary findings remain consistent after adjusting for severity of airflow obstruction.

Pathway analyses

Significant gene ontology enrichment results were selected with topGO weight01 statistics P-value < 0.001, as a stringent criterion for statistical significance. Gene set enrichment analyses results were selected with FDR controlled at 0.05.

Weighted gene correlation network analyses

The selection of the soft-thresholding power (β) is a critical step for constructing a biologically meaningful network. The soft-thresholding power controls the extent to which strong gene-gene correlations are emphasised while weak correlations are suppressed. We set minimal module size at ten and optimised soft threshold (power) according to scale-free topology model fit and mean connectivity to evaluate connectivity between network nodes.

Phenotypic and proteomic genetic associations

Population stratification covariates were added for linear and logistic regression models. Significant phenotypic genetic associations were selected with P-value < 0.05 and significant quantitative trait loci associations were selected with FDR controlled at 0.05.

Machine learning modelling

With ntree (number of decision trees) tuning from 400 to 1000 and mtry (number of variables sampled for each decision tree) tuning from 3 to 12, we utilised 10-fold cross-validation to tune parameters: number of trees and mtry with optimised prediction accuracy. A feature selection approach was performed for each training process using the mutual information between each protein and COPD subtypes. The number of optimised features was rounded to the square root of training set sample number.

Role of funders

The study sponsors had no role in the study design; in the collection, analysis, and interpretation of the data; in the writing of the report; and in the decision to submit the paper for publication.

Results

Clinical characteristics

Participants were classified into control, EPD, and NEPD groups based on spirometry and quantitative CT emphysema. Clinical and demographic characteristics are presented in Table 1 (COPDGene) and Supplemental Table S1 (SPIROMICS). A longer smoking history and worse lung function values were observed in EPD compared to participants with NEPD. The CONSORT diagram for COPD subtype proteomics analysis is shown in Fig. 1.

Table 1.

Clinical characteristics of control, and participants with emphysema-predominant COPD, and non-emphysema predominant COPD in the COPDGene cohort.

	Controlb,c	Non-emphysema predominant COPDb,c	Emphysema-predominant COPDb,c	P for emphysema/non-emphysema predominant COPDb
Number of participants	2091	722	726
Age (Years)	63.11 (8.28)	65.84 (8.34)	69.15 (7.71)	<0.0001
Sex (Male/All subjects)	0.46 (952/2091)	0.49 (352/722)	0.59 (426/726)	0.0002
Sex (Female/All subjects)	0.54 (1139/2091)	0.51 (370/722)	0.41 (300/726)	0.0002
Race (White/All subjects)e	0.65 (1364/2091)	0.73 (524/722)	0.79 (570/726)	0.01
Pack-yearsd	34.80 (23.80)	44.00 (26.20)	48.00 (31.65)	<0.0001
BMI (Kg/M2)a	29.36 (6.00)	30.38 (6.68)	25.94 (5.17)	<0.0001
Smoking status (Former smokers/All subjects)	0.59 (1235/2091)	0.51 (366/722)	0.79 (574/726)	<0.0001
FEV1/FVC Ratioa	0.78 (0.05)	0.60 (0.08)	0.43 (0.10)	<0.0001
FEV1% predicted	97.54 (11.58)	61.12 (12.84)	43.29 (16.54)	<0.0001
PRM fSAD (%)a,d	6.93 (6.58)	16.26 (12.91)	34.56 (14.84)	<0.0001
Chronic bronchitisa (Fraction of affirmative responses)	0.09 (185/2091)	0.24 (173/722)	0.26 (185/726)	0.5
Six-minute walk distance (Metres)	433.97 (120.30)	365.03 (122.35)	320.66 (135.04)	<0.0001
Pi10a	1.99 (0.41)	2.74 (0.627)	2.66 (0.49)	0.004
DLCO (% predicted)a	88.70 (17.55)	74.63 (18.46)	48.63 (17.02)	<0.0001
Total corticosteroid usea (User/All subjects)	0.03 (59/2091)	0.08 (60/722)	0.13 (93/726)	0.007

^aBMI: body mass index; PRM fSAD: Parametric Response Mapping, functional small airway disease; Pi10: square root of wall area for a hypothetical airway of 10-mm lumen perimeter on computed tomography; FEV1: amount of air forcefully exhaled in the first second after taking a deep breath; FVC: the total amount of air forcefully exhaled after a full inhalation; DLCO: the diffusing capacity of the lungs for carbon monoxide; Total corticosteroid use: inhaled and/or oral corticosteroids use reported at the COPDGene Phase 2 visit.

^bChi square tests were used for categorical variables and T tests were used for continuous variables.

^cMean (SD) for quantitative variables without skewness and proportion (n/N) for categorical variables.

^dMedian (IQR) for variables with skewness: PRM fSAD (%) and Pack-years.

^eParticipants were enrolled in two racial groups: non-Hispanic White and Black.

Proteomic biomarker identification

Power calculations confirming adequate sample size for proteomic biomarker detection are presented in Supplemental Figure S1. PCA plots for proteomics residuals in the COPDGene and SPIROMICS cohorts, which demonstrate overlap between participants with EPD and participants with NEPD for most proteins, are presented in Supplemental Figure S2. We identified 125 SOMAmers representing 124 proteins (UniProt ID) at FDR < 0.05 comparing EPD and NEPD COPD within COPDGene, as shown in Supplemental Table S2. sRAGE (encoded by AGER) is the most significant subtype-associated proteomic biomarker identified in COPDGene. We visualised the directions and significance of COPD subtype proteomic associations in COPDGene and SPIROMICS using volcano plots, as shown in Fig. 2 (COPDGene cohort) and Supplemental Figure S3 (SPIROMICS). The top 20 proteomic biomarkers in COPDGene (FDR < 0.05) that were at least nominally associated with COPD subtypes in SPIROMICS (P-value < 0.05) are shown in Table 2; the full list of 66 COPDGene SomaScan biomarkers (SOMAmers, representing 65 UniProt IDs) with nominal P-value < 0.05 in SPIROMICS is shown in Supplemental Table S3. Only one of these 66 associations (STOML1) had an opposite direction of effect in COPDGene and SPIROMICS. Of interest, five proteins were associated with COPD subtypes with FDR < 0.05 in both COPDGene and SPIROMICS (sRAGE, KRT20, CD93, KLK10, and WFDC1). Among 124 COPD subtype proteomic biomarkers (UniProt ID) in COPDGene, 120 had the same direction of effect in COPDGene and SPIROMICS, a highly statistically significant concordance in effect direction (P = 9.12 × 10⁻³¹, as shown in Supplemental Table S5). For all 4727 proteins (UniProt ID) assessed in COPDGene, 59% had higher levels in participants with EPD; however, among the 124 subtype proteomic biomarkers, 88% had higher protein levels in participants with NEPD. A similar pattern was observed in SPIROMICS.

Fig. 2.

Volcano plot for proteomic biomarkers identified by comparison between emphysema-predominant and non-emphysema predominant COPD. Models were adjusted for age, sex, race, pack-years of smoking, BMI, current smoking status, clinical centre, and cellular components (white blood cell counts and platelet counts).

Table 2.

Top twenty emphysema-predominant vs. non-emphysema predominant COPD proteomic biomarkers in COPDGene with nominal replication (P < 0.05) in SPIROMICS, ranking by P-values in COPDGene.

Protein name	Gene name	UniProt ID	SOMAmer	COPDGene P	COPDGene FDR	COPDGene CIs	SPIROMICS P	SPIROMICS CIs
Advanced glycosylation end product-specific receptor, soluble	AGER	Q15109	4125_52	3.2E-15	1.6E-11	−0.36, −0.22	2.8E-05	−0.45, −0.16
N-terminal pro-BNP	NPPB	P16860	7655_11	7.0E-09	1.3E-05	−0.38, −0.19	3.0E-02	−0.34, −0.02
Polymeric immunoglobulin receptor	PIGR	P01833	3216_2	1.2E-08	1.5E-05	−0.22, −0.11	1.1E-03	−0.30, −0.08
Collagen alpha-1(XXVIII) chain	COL28A1	Q2UY09	10702_1	6.1E-08	6.0E-05	−0.13, −0.06	2.2E-03	−0.15, −0.03
Peptidyl-prolyl cis-trans isomerase C	PPIC	P45877	18819_21	7.5E-08	6.2E-05	−0.12, −0.06	4.6E-04	−0.16, −0.05
Laminin subunit gamma 2	LAMC2	Q13753	9580_5	2.7E-07	1.9E-04	−0.16, −0.07	2.1E-03	−0.21, −0.05
EGF-like repeat and discoidin I-like domain-containing protein 3	EDIL3	O43854	9360_33	5.8E-07	3.6E-04	−0.17, −0.08	2.2E-03	−0.24, −0.06
Peroxidasin homologue	PXDN	Q92626	13463_1	7.6E-07	3.8E-04	−0.2, −0.09	3.3E-02	−0.21, −0.01
Neuroblastoma suppressor of tumorigenicity 1	NBL1	P41271	2944_66	1.6E-06	6.5E-04	−0.16, −0.07	2.8E-03	−0.16, −0.03
Trefoil factor 2	TFF2	Q03403	9191_8	2.2E-06	8.2E-04	−0.18, −0.07	2.3E-03	−0.26, −0.06
Complement component C1q receptor	CD93	Q9NPY3	14136_234	3.2E-06	1.1E-03	−0.09, −0.04	2.0E-08	−0.20, −0.10
Collagen alpha-3(VI) chain	COL6A3	P12111	11196_31	3.8E-06	1.3E-03	−0.11, −0.04	3.4E-04	−0.16, −0.05
Secreted frizzled-related protein 4	SFRP4	Q6FHJ7	17447_52	5.2E-06	1.5E-03	−0.13, −0.05	6.8E-04	−0.20, −0.05
WAP four-disulfide core domain protein 1	WFDC1	Q9HC57	9316_67	5.6E-06	1.5E-03	−0.12, −0.05	5.3E-05	−0.21, −0.07
Ephrin-B2: extracellular domain	EFNB2	P52799	14131_37	5.8E-06	1.5E-03	−0.14, −0.06	1.9E-04	−0.23, −0.07
Ephrin-A4	EFNA4	P52798	2614_28	8.5E-06	2.0E-03	−0.09, −0.04	3.9E-03	−0.12, −0.03
Desmocollin-2	DSC2	Q02487	13126_52	1.0E-05	2.1E-03	−0.10, −0.04	1.5E-02	−0.15, −0.02
Complement decay-accelerating factor	CD55	P08174	5069_9	1.7E-05	3.4E-03	−0.08, −0.03	6.6E-04	−0.12, −0.03
Netrin receptor UNC5B	UNC5B	Q8IZJ1	15394_79	1.9E-05	3.5E-03	−0.12, −0.05	5.4E-04	−0.21, −0.06
Inactive tyrosine-protein kinase transmembrane receptor ROR1	ROR1	Q01973	2590_69	2.4E-05	4.2E-03	−0.12, −0.04	7.4E-03	−0.18, −0.03

P (P-value), FDR (false discovery rate), and CIs (Confidence Intervals for Beta coefficient from linear regression model).

As an alternative approach to identify COPD subtype proteomic biomarkers that leverages the large number of control participants in COPDGene, we evaluated the proteomic expression differences among controls, participants with EPD, and participants with NEPD using ANOVA (one-way ANOVA) and post-hoc SNK tests.⁴¹ Most of the proteomic target biomarkers (102/124 at UniProt ID level, 103/125 at SomaScan Probe level) identified in the linear regression models were also significant using the ANOVA/SNK approach, as shown in Supplemental Figure S4. Comparison of protein expression levels for COPD subtype biomarkers in control, emphysema-predominant COPD, and non-emphysema predominant COPD are shown in Supplemental Table S7: most proteomic biomarkers had similar levels in participants with NEPD and control, and significantly lower values in participants with EPD.

To evaluate the impact of airflow obstruction severity on proteomic biomarker detection, we identified significant COPD subtype proteomic associations with adjustment for FEV1% predicted (FDR controlled at 0.05) in Supplemental Table S4. Fifty-nine of the original 124 protein biomarkers (two biomarkers with higher expression level in EPD and 57 biomarkers with higher expression level in NEPD) were consistent after adjusting for FEV1% predicted with FDR controlled at 0.05, as shown in Supplemental Figure S5.

Proteomic expression levels for two previously reported proteins in COPD pathogenesis, sRAGE²¹ and PIGR,^42,43 are shown in Supplemental Figure S6. Lower protein levels were observed in EPD for both proteins compared to participants with NEPD and control.

Pair-wise protein-protein correlations were evaluated in COPD subtype proteomic biomarkers. Higher correlations were observed between biomarkers with higher expression levels in participants with NEPD, as shown in Supplemental Figure S7.

We also compared identified COPD subtype biomarkers with previously reported cardiac biomarkers (60 detectable in COPDGene SOMAScan proteomics panel) and found five overlapping biomarkers including NPPB,⁴⁴ DSC2,⁴⁵ CST3,⁴⁶ KRT1,⁴⁷ and ADH1B.⁴⁸ All five proteins have higher expression levels in non-emphysema predominant COPD and all were previously reported to be expressed more highly in cardiac diseases (Supplemental Table S8).

Pathway analysis

Since the vast majority of the 124 protein biomarkers identified in COPDGene had a similar direction of effect in SPIROMICS, we performed gene set enrichment analysis and gene ontology enrichment analysis using the 124 significant protein biomarkers from COPDGene. Significant results for HALLMARK gene sets and GO enrichment are shown in Fig. 3. Significant KEGG and REACTOME gene set enrichment results are presented in Supplemental Figure S8. COPD subtype-associated proteomic biomarkers are significantly enriched for important biological processes including cell adhesion, collagen-containing extracellular matrix, and epithelial mesenchymal transition.

Fig. 3.

HALLMARK Gene sets and gene ontology enrichment analysis for COPD subtype proteomic associations. Significant gene sets were selected with q-value < 0.05, and significant gene ontology terms were selected with P-value < 0.001.

Weighted gene correlation network analysis

The power threshold for WGCNA was optimised as shown in Supplemental Figure S9. With minimum network module size controlled at 10, we identified eleven correlation network modules. Only the green module was associated with EPD vs. NEPD COPD subtypes with FDR controlled at 0.05. COPD subtype proteomic association P-values across different network modules are compared in Supplemental Figure S10. Proteins in the green module were more significantly associated with COPD subtypes compared to other modules. There are 105 proteins (represented by 108 SOMAmers) in module green and among them, 40 proteins (42 SOMAmers) are COPD subtype-associated proteomic biomarkers. Module membership is the correlation between module eigengene and module components and reflects the intramodular connectivity of each component.³⁸ Gene significance (COPD subtype associations) and module membership for green module components are listed in Supplemental Table S9 and their associations are shown in Supplemental Figure S11. With COPD subtype association FDR controlled at 0.05 and module membership > 0.8, we identified thirteen COPD subtype protein biomarkers representative for the green module. These highly connected protein biomarkers were PXDN, EFNA2, EFNB2, SELENOM, UNC5B, EPHB4, LRP10, ROR1, TNFRSF1A, EPHA2, GAS1, ROR2, and SPON2.

Gene ontology enrichment analysis was performed with green module components as proteins of interest and all detected human proteins as background. Proteins from the green COPD subtype-associated module enrich in biological processes including cell adhesion and ephrin receptor signalling pathway with TopGO Fisher's P-value < 0.001, as shown in Supplemental Figure S12.

Genetic associations for COPD subtypes and COPD subtype proteomic biomarkers

We analysed top SNPs from 82 COPD GWAS loci²⁷ and rs28929474 in SERPINA1, as shown in Table 3. Seven nominally significant (P < 0.05) EPD vs. NEPD associations for COPD genetic loci were found including rs28929474 in SERPINA1; the only COPD-associated SNP significant at FDR < 0.05 was rs2579762 in LRMDA as shown in Supplemental Table S10. Associations of COPD SNPs with each COPD subtype/control comparison are presented in Supplemental Figure S13. SNPs near ADGRG6 and TGFB2 were nominally associated with EPD vs. Control only, while another SNP near STN1 was associated with NEPD vs. Control only.

Table 3.

Genetic variant information for 82 GWAS loci reported in Sakornsakolpat et al. and rs28929474 in SERPINA1.

RS ID	Chromosome	Position	Risk/alternative allele	Nearest gene	COPD odds ratio	COPD GWAS P-value
rs13140176	chr4	145489098	A/G	HHIP	1.18	4.1E-59
rs34712979	chr4	106819053	A/G	NPNT	1.18	3.0E-46
rs9399401	chr6	142668901	T/C	ADGRG6	1.16	1.6E-40
rs10037493	chr5	147854970	C/T	HTR4	1.13	2.6E-33
rs1441358	chr15	71612514	G/T	THSD4	1.13	7.4E-33
rs55676755	chr15	78898932	G/C	CHRNA3	1.11	2.7E-26
rs7068966	chr10	12277992	C/T	CDC123	1.1	6.2E-23
rs3095329	chr6	30693816	G/A	IER3	1.12	2.1E-21
rs4888379	chr16	75340231	T/A	CFDP1	1.1	5.9E-21
rs16825267	chr2	229569919	C/G	PID1	1.19	1.8E-20
rs7671261	chr4	89883818	A/G	FAM13A	1.09	1.4E-18
rs2070600	chr6	32151443	C/T	AGER	1.21	1.1E-17
rs10866659	chr5	156937043	G/A	ADAM19	1.09	1.2E-16
rs2806356	chr6	109266255	C/T	ARMC2	1.1	2.9E-15
rs2955083	chr3	127961178	A/T	EEFSEC	1.13	3.5E-15
rs7642001	chr3	168746145	A/G	MECOM	1.08	1.9E-14
rs9350191	chr6	19842661	T/C	ID4	1.12	5.1E-14
rs10114763	chr9	4143749	T/A	GLIS3	1.07	8.7E-13
rs62191105	chr2	239872704	C/T	TWIST2	1.09	2.6E-12
rs2571445	chr2	218683154	A/G	TNS1	1.07	3.4E-12
rs156394	chr9	23588684	T/C	ELAVL2	1.07	3.8E-12
rs10152300	chr15	84392907	G/A	ADAMTSL3	1.08	4.2E-12
rs647097	chr18	8808464	C/T	MTCL1	1.08	1.0E-11
rs1529672	chr3	25520582	C/A	RARB	1.09	2.5E-11
rs11118406	chr1	219924894	T/A	SLC30A10	1.08	4.1E-11
rs646695	chr6	140280398	C/T	CITED2	1.08	4.6E-11
rs17759204	chr3	55158224	G/A	CACNA2D3	1.07	8.8E-11
rs10760580	chr9	101661650	G/A	COL15A1	1.07	1.2E-10
rs955277	chr2	9290357	T/C	ASAP2	1.07	1.9E-10
rs2442776	chr3	11640601	G/A	VGLL4	1.09	2.0E-10
rs2579762	chr10	78318879	C/A	LRMDA	1.06	2.6E-10
rs629619	chr1	111738108	T/C	DENND2D	1.08	2.9E-10
rs9329170	chr8	8697658	C/G	MFHAS1	1.1	3.6E-10
rs4093840	chr3	123077042	A/T	ADCY5	1.06	3.9E-10
rs9617650	chr22	18488883	G/C	MICAL3	1.08	4.4E-10
rs76841360	chr1	40060025	A/G	PABPC4	1.08	5.0E-10
rs1551943	chr5	52195033	A/G	ITGA1	1.08	5.8E-10
rs153916	chr5	95036700	T/C	SPATA9	1.06	6.3E-10
rs9435731	chr1	17306029	A/C	MFAP2	1.06	6.6E-10
rs117261012	chr11	86444761	G/A	PRSS23	1.09	6.9E-10
rs2897075	chr7	99630342	C/T	ZKSCAN1	1.07	7.3E-10
rs12185268	chr17	43923683	G/A	SPPL2C	1.08	9.9E-10
rs7958945	chr12	115947901	G/A	MED13L	1.06	1.0E-09
rs12519165	chr5	170901586	A/T	FGF18	1.07	1.1E-09
rs13198656	chr6	22004909	T/C	PRL	1.06	1.2E-09
rs72626215	chr19	46294136	G/A	DMWD	1.07	1.7E-09
rs11655567	chr17	69216687	C/T	SOX9	1.06	1.9E-09
rs7307510	chr12	96237570	C/T	SNRPF	1.08	2.6E-09
rs9525927	chr13	44842503	G/A	SERP2	1.08	2.8E-09
rs4585380	chr4	75673363	G/A	BTC	1.07	3.4E-09
rs4757118	chr11	13171236	T/C	ARNTL	1.06	3.8E-09
rs798565	chr7	2752152	G/A	AMZ1	1.07	3.9E-09
rs72673419	chr1	60913143	T/C	C1orf87	1.14	4.0E-09
rs3009947	chr1	218689155	C/T	TGFB2	1.06	4.0E-09
rs72699855	chr14	93105953	G/C	RIN3	1.08	4.8E-09
rs11579382	chr1	239901006	C/G	CHRM3	1.06	6.5E-09
rs2040732	chr7	20418134	C/T	ITGB8	1.06	6.9E-09
rs1631199	chr6	117259673	G/C	RFX6	1.06	7.6E-09
rs73158393	chr22	33335386	C/G	SYN3	1.07	7.7E-09
rs62065216	chr17	38218773	A/G	THRA	1.06	8.1E-09
rs72731149	chr15	49984710	G/C	DTWD1	1.12	8.3E-09
rs10929386	chr2	15906179	C/T	DDX1	1.06	9.1E-09
rs34727469	chr17	36835079	T/C	RPL23	1.09	1.1E-08
rs8044657	chr16	58022625	G/A	TEPP	1.11	1.1E-08
rs1334576	chr6	7211818	A/G	RREB1	1.06	1.2E-08
rs8080772	chr17	28413129	T/C	EFCAB5	1.06	1.4E-08
rs979453	chr5	150595073	G/A	CCDC69	1.06	1.4E-08
rs72902175	chr2	157013035	T/C	NR4A2	1.09	1.6E-08
rs7866939	chr9	85126163	C/T	RASEF	1.06	1.8E-08
rs13073544	chr3	29472412	C/G	RBMS3	1.06	2.0E-08
rs721917	chr10	81706324	G/A	SFTPD	1.06	2.2E-08
rs62375246	chr5	132439010	A/T	HSPA4	1.06	2.2E-08
rs1570221	chr10	105656874	A/G	STN1	1.06	2.2E-08
rs62259026	chr3	57746515	C/T	SLMAP	1.07	2.4E-08
rs11049386	chr12	28320536	T/A	CCDC91	1.06	2.7E-08
rs803923	chr9	119401650	A/G	ASTN2	1.06	2.7E-08
rs34651	chr5	72144005	C/T	TNPO1	1.11	3.0E-08
rs2096468	chr21	35661745	A/C	KCNE2	1.06	4.0E-08
rs4660861	chr1	45946636	G/T	TESK2	1.06	4.4E-08
rs56134392	chr16	10709013	C/T	TEKT5	1.06	4.5E-08
rs12466981	chr2	42433247	C/T	EML4	1.06	4.9E-08
rs7650602	chr3	141147414	C/T	ZBTB38	1.06	4.9E-08
rs28929474	chr14	94844947	T/C	SERPINA1	Not a GWAS Variant

Protein information for the COPD subtype proteomic biomarkers in COPDGene used for pQTL analysis is shown in Table 4. QQ plots for cis (±1 Mb with 83 COPD SNPs) and trans (all SNPs beyond ±1 Mb from a protein biomarker gene) pQTL analysis associating COPD SNPs with COPD subtype proteomic biomarkers are shown in Supplemental Figure S14. Power calculations confirming adequate sample size using R package pwr for QTL analysis are presented in Supplemental Figure S1. No significant trans-pQTL associations were found. We identified significant cis-pQTL associations with FDR controlled at 0.05 between rs2070600 and sRAGE in all participants, and separately in participants who are controls, participants with EPD, and participants with NEPD, and between rs979453 and GM2A in all participants and separately in control and NEPD populations, as shown in Supplemental Table S11.

Table 4.

Protein information (GRCh37) for COPD subtype associated proteomic biomarkers in COPDGene cohort for pQTL analysis.

Protein name	Gene name	UniProt ID	SOMAmer	FDR	Chr	Start	End
Advanced glycosylation end product-specific receptor, soluble	AGER	Q15109	4125_52	1.6E-11	6	32148745	32152101
Glucagon	GCG	P01275	4891_50	1.3E-05	2	162999392	163008914
N-terminal pro-BNP	NPPB	P16860	7655_11	1.3E-05	1	11917521	11918988
Polymeric immunoglobulin receptor	PIGR	P01833	3216_2	1.5E-05	1	207101863	207119811
Collagen alpha-1(XXVIII) chain	COL28A1	Q2UY09	10702_1	6.0E-05	7	7395834	7575484
Peptidyl-prolyl cis-trans isomerase C	PPIC	P45877	18819_21	6.2E-05	5	122358945	122372436
Laminin subunit gamma 2	LAMC2	Q13753	9580_5	1.9E-04	1	183155373	183214035
EGF-like repeat and discoidin I-like domain-containing protein 3	EDIL3	O43854	9360_33	3.6E-04	5	83236373	83680611
Peroxidasin homologue	PXDN	Q92626	13463_1	3.8E-04	2	1635659	1748624
Neurotensin/neuromedin N	NTS	P30990	7857_22	3.8E-04	12	86268073	86276770
N-acetylglucosamine-1-phosphotransferase subunit gamma	GNPTG	Q9UJJ9	10666_7	3.8E-04	16	1401924	1413352
Neuroblastoma suppressor of tumorigenicity 1	NBL1	P41271	2944_66	6.5E-04	1	19967048	19984945
Trefoil factor 2	TFF2	Q03403	9191_8	8.2E-04	21	43766466	43771237
Complement component C1q receptor	CD93	Q9NPY3	14136_234	1.1E-03	20	23059986	23066977
Collagen alpha-3(VI) chain 1	COL6A3	P12111	11196_31	1.3E-03	2	238232646	238323018
RELT-like protein 1	RELL1	Q8IUW5	13399_33	1.4E-03	4	37592422	37687998
Ephrin-B2: extracellular domain	EFNB2	P52799	14131_37	1.5E-03	13	107142079	107187462
Secreted frizzled-related protein 4	SFRP4	Q6FHJ7	17447_52	1.5E-03	7	37945543	38065297
WAP four-disulfide core domain protein 1	WFDC1	Q9HC57	9316_67	1.5E-03	16	84328252	84363450
Retinol-binding protein 4	RBP4	P02753	15633_6	2.0E-03	10	95351444	95361501
Ephrin-A4	EFNA4	P52798	2614_28	2.0E-03	1	155036207	155042029
Peptide YY	PYY	P10082	3727_35	2.1E-03	17	42030106	42081837
Soluble calcium-activated nucleotidase 1	CANT1	Q8WVQ1	6480_1	2.1E-03	17	76987799	77005949
Desmocollin-2	DSC2	Q02487	13126_52	2.1E-03	18	28645940	28682378
Complement decay-accelerating factor	CD55	P08174	5069_9	3.4E-03	1	207494853	207534311
Netrin receptor UNC5B	UNC5B	Q8IZJ1	15394_79	3.5E-03	10	72972327	73062621
Follistatin-related protein 3	FSTL3	O95633	3438_10	3.5E-03	19	676392	683385
Inactive tyrosine-protein kinase transmembrane receptor ROR1	ROR1	Q01973	2590_69	4.2E-03	1	64239693	64647181
Ganglioside GM2 activator	GM2A	P17900	15441_6	5.1E-03	5	150591711	150650001
Ribonuclease T2	RNASET2	O00584	16913_8	5.1E-03	6	167342992	167370679
Neurocan core protein	NCAN	O14594	15573_110	5.5E-03	19	19322782	19363042
Ephrin type-A receptor 2	EPHA2	P29317	4834_61	5.9E-03	1	16450832	16482582
Tyrosine-protein kinase transmembrane receptor ROR2	ROR2	Q01974	7861_9	6.0E-03	9	94325373	94712444
Beta-2-microglobulin	B2M	P61769	3485_28	6.3E-03	15	45003675	45011075
Myocilin	MYOC	Q99972	16558_2	6.6E-03	1	171604557	171621823
Xyloside xylosyltransferase 1	XXYLT1	Q8NBI6	6375_75	6.6E-03	3	194789008	194991896
Tryptase beta 1	TPSAB1	Q15661	9409_11	6.6E-03	16	1290697	1292555
Interleukin-18	IL18	Q14116	5661_15	6.8E-03	11	112013974	112034840
Thrombospondin-2	THBS2	P35442	3339_33	7.2E-03	6	169615875	169654139
Vesicular integral-membrane protein VIP36	LMAN2	Q12907	9468_8	7.5E-03	5	176758563	176778853
Retinol-binding protein 5	RBP5	P82980	19241_31	8.4E-03	12	7276280	7281538
Prefoldin subunit 5	PFDN5	Q99471	4271_75	8.4E-03	12	53689075	53693234
Kallikrein-10	KLK10	O43240	6227_1	8.6E-03	19	51515995	51523431
ADAMTS-like protein 2	ADAMTSL2	Q86TH1	6379_62	9.1E-03	9	136397286	136440641
Tumour necrosis factor receptor superfamily member 1A	TNFRSF1A	P19438	2654_19	1.1E-02	12	6437923	6451280
C-type mannose receptor 2	MRC2	Q9UBG0	3041_55	1.1E-02	17	60704762	60770958
Lumican	LUM	P51884	13114_50	1.1E-02	12	91496406	91505608
Brain-specific serine protease 4	PRSS22	Q9GZN4	4534_10	1.1E-02	16	2902728	2908171
Matrilin-4	MATN4	O95460	7083_74	1.1E-02	20	43922085	43937169
Ribonuclease UK114	RIDA	P52758	14636_25	1.2E-02	8	99114572	99129469
Protein S100-A16	S100A16	Q96FQ6	17836_17	1.2E-02	1	153579362	153585621
Out at first protein homologue	OAF	Q86UD1	6414_8	1.2E-02	11	120081475	120101041
Gastrokine-2	GKN2	Q86XP6	6416_8	1.2E-02	2	69172364	69180102
Left-right determination factor 2	LEFTY2	O00292	15503_15	1.3E-02	1	226124298	226129189
Cadherin-1	CDH1	P12830	18429_10	1.3E-02	16	68771128	68869451
Dual specificity mitogen-activated protein kinase kinase 2	MAP2K2	P36507	3628_3	1.3E-02	19	4090319	4124126
CMRF35-like molecule 6	CD300C	Q08708	5066_134	1.3E-02	17	72537247	72542282
Alkaline phosphatase, placental-like	ALPG	P10696	6715_63	1.3E-02	2	233271553	233275424
Neuronal pentraxin-1	NPTX1	Q15818	9256_78	1.3E-02	17	78440948	78451643
Cadherin-17	CDH17	Q12864	16613_3	1.4E-02	8	95139399	95229531
Growth arrest-specific protein 1	GAS1	P54826	5463_22	1.4E-02	9	89559279	89562104
Natural cytotoxicity triggering receptor 1	NCR1	O76036	8360_169	1.5E-02	19	55417508	55427508
High mobility group protein B1	HMGB1	P09429	2524_56	1.6E-02	13	31032884	31191734
Tyrosine-protein phosphatase non-receptor type 1	PTPN1	P18031	3005_5	1.6E-02	20	49126891	49201299
Ribonuclease pancreatic	RNASE1	P07998	7211_2	1.6E-02	14	21269387	21271437
Hepatitis A virus cellular receptor 2	HAVCR2	Q8TDQ0	5134_52	1.7E-02	5	156512843	156569880
Leucocyte immunoglobulin-like receptor subfamily A member 3	LILRA3	Q8N6C8	6391_52	1.8E-02	19	54799854	54809952
Netrin receptor UNC5B	UNC5B	Q8IZJ1	7776_20	1.8E-02	10	72972327	73062621
Interleukin-2 receptor subunit beta	IL2RB	P14784	9343_16	1.8E-02	22	37521878	37571094
Cyclin-dependent kinase 8: Cyclin-C complex	CCNC	P24863	3359_11	2.1E-02	6	99990256	100016849
	CDK8	P49336			13	26828276	26979375
Ankyrin repeat and SOCS box protein 9	ASB9	Q96DX5	19601_15	2.1E-02	X	15253410	15288589
Amyloid-like protein 1	APLP1	P51693	7210_25	2.1E-02	19	36358801	36370693
Tumour necrosis factor receptor superfamily member 1B	TNFRSF1B	P20333	8368_102	2.1E-02	1	12227060	12269285
Cartilage oligomeric matrix protein	COMP	P49747	8043_153	2.1E-02	19	18893583	18902123
Selenoprotein M	SELENOM	Q8WWX9	15336_7	2.2E-02	22	31500758	31516055
Neural cell adhesion molecule 1120 kDa isoform	NCAM1	P13591	4498_62	2.3E-02	11	112831997	113149158
Carbonyl reductase [NADPH] 1	CBR1	P16152	12381_26	2.3E-02	21	37442239	37445464
Microfibril-associated glycoprotein 4	MFAP4	P55083	5636_10	2.3E-02	17	19286755	19290553
Cystatin-C	CST3	P01034	2609_59	2.4E-02	20	23608534	23619110
Pigment epithelium-derived factor	SERPINF1	P36955	9211_19	2.4E-02	17	1665253	1680868
Ephrin-A2	EFNA2	O43921	14124_6	2.5E-02	19	1286153	1301430
Glutamate carboxypeptidase 2	FOLH1	Q04609	5478_50	2.5E-02	11	49168187	49230222
Aldo-keto reductase family 1 member C3	AKR1C3	P42330	17377_1	2.6E-02	10	5077546	5149878
Thrombospondin-4	THBS4	P35443	3340_53	2.8E-02	5	79287134	79379110
Matrix-remodelling-associated protein 7	MXRA7	P84157	8005_1	2.8E-02	17	74668633	74707098
SPARC-related modular calcium-binding protein 1	SMOC1	Q9H4F8	13118_5	2.8E-02	14	70320848	70499083
SPARC-related modular calcium-binding protein 2	SMOC2	Q9H3U7	15635_4	2.8E-02	6	168841831	169073984
Low-density lipoprotein receptor-related protein 10	LRP10	Q7Z4F1	16610_13	2.8E-02	14	23340822	23350789
Tryptase beta 2	TPSB2	P20231	3403_1	2.8E-02	16	1290697	1292555
IGF-like family receptor 1	IGFLR1	Q9H665	7244_16	2.8E-02	19	36230058	36233354
Keratin, type II cytoskeletal 1	KRT1	P04264	9931_20	2.8E-02	12	53068520	53074191
Netrin-G1	NTNG1	Q9Y2I2	5637_81	2.8E-02	1	107682629	108026080
Ephrin type-B receptor 4	EPHB4	P54760	15530_33	2.9E-02	7	100400187	100425121
Thrombospondin-3	THBS3	P49746	8982_65	2.9E-02	1	155165379	155178842
Keratin, type I cytoskeletal 20	KRT20	P35900	12975_11	3.2E-02	17	39032193	39041479
Cathepsin E	CTSE	P14091	15376_134	3.2E-02	1	206317459	206332104
Matrilin-2	MATN2	O00339	3325_2	3.2E-02	8	98881068	99048944
Procollagen galactosyltransferase 1	COLGALT1	Q8NBJ5	5638_23	3.2E-02	19	17666403	17693971
EGF-containing fibulin-like extracellular matrix protein 1	EFEMP1	Q12805	8480_29	3.2E-02	2	56093102	56151274
Polypeptide N-acetylgalacto-saminyltransferase 16	GALNT16	Q8N428	8923_94	3.2E-02	14	69725994	69821183
Alcohol dehydrogenase 1B	ADH1B	P00325	9834_62	3.2E-02	4	100226121	100242558
Tyrosine-protein phosphatase non-receptor type 4	PTPN4	P29074	14254_27	3.3E-02	2	120517207	120741394
RGM domain family member B	RGMB	Q6NW40	3331_8	3.3E-02	5	98104354	98134347
Follistatin	FST	P19883	4132_27	3.3E-02	5	52776239	52782964
Bone morphogenetic protein 7	BMP7	P18075	2972_57	3.4E-02	20	55743804	55841685
Tumour necrosis factor receptor superfamily member 1B	TNFRSF1B	P20333	3152_57	3.4E-02	1	12227060	12269285
Serine protease inhibitor Kazal-type 9	SPINK9	Q5DT21	8042_88	3.4E-02	5	147700766	147719412
Stomatin-like protein 1	STOML1	Q9UBI4	17344_23	3.5E-02	15	74275547	74286963
Interleukin-18 receptor accessory protein	IL18RAP	O95256	2993_1	3.6E-02	2	103035149	103069025
Apolipoprotein L1	APOL1	O14791	9506_10	3.6E-02	22	36649056	36663576
Vascular endothelial growth factor D	VEGFD	O43915	13098_93	3.6E-02	X	15363713	15402498
Cell adhesion molecule 1	CADM1	Q9BY67	3326_58	3.6E-02	11	115039938	115375675
Alpha-1-microglobulin	AMBP	P02760	15453_3	3.6E-02	9	116822407	116840752
CD48 antigen	CD48	P09326	3292_75	3.6E-02	1	160648536	160681641
Membrane cofactor protein	CD46	P15529	17682_1	3.9E-02	1	207925402	207968858
Cathepsin Z	CTSZ	Q9UBR2	4971_1	3.9E-02	20	57570240	57582302
Voltage-dependent calcium channel subunit alpha-2/delta-3	CACNA2D3	Q8IZS8	8885_6	4.4E-02	3	54156574	55108584
Brorin	VWC2	Q2TAL6	15308_108	4.6E-02	7	49813257	49961546
Glycine N-acyltransferase	GLYAT	Q6IB77	19506_6	4.6E-02	11	58407899	58499447
Calpain-2 catalytic subunit	CAPN2	P17655	14684_17	4.8E-02	1	223889295	223963720
Cellular retinoic acid-binding protein 2	CRABP2	P29373	11696_7	<5.0E-02	1	156669398	156675608
Low-density lipoprotein receptor-related protein 11	LRP11	Q86VZ4	15472_16	<5.0E-02	6	150139934	150186199
Ribonuclease K6	RNASE6	Q93091	5646_20	<5.0E-02	14	21249210	21250626
Complement C1q tumour necrosis factor-related protein 5	C1QTNF5	Q9BXJ0	7810_20	<5.0E-02	11	119209652	119217383
Spondin-2	SPON2	Q9BUD6	8099_42	<5.0E-02	4	1160720	1202750

Machine learning model for COPD subtype classification

We established a random forest model for COPD subtype classification in COPDGene. We optimised the random forest model using classification accuracy and identified the optimised random forest model with number of trees as 600 and mtry as 8. The final model was further evaluated in SPIROMICS with prediction accuracy as 0.60 and area under curve (AUC) as 0.65 (CIs: 0.61–0.69), as shown in Supplemental Figure S15. Top optimised features selected by the machine learning model are presented in Supplemental Figure S16, including significant EPD/NEPD proteomic biomarkers AGER and PIGR.

Additional information regarding Results is provided in the Supplemental Materials.

Discussion

Participants with emphysema predominant and non-emphysema predominant COPD have differences in clinical characteristics and disease progression. Castaldi et al. recently demonstrated that multiple different approaches for COPD subtyping could be well-captured using the EPD/NEPD criteria that we adopt in these analyses.¹² However, the biological mechanisms distinguishing EPD and NEPD COPD subtypes are unknown. Our analysis focused on identification of proteomic biomarkers, protein correlations, biomarker-enriched biological pathways, correlation networks, and genetic regulatory effects to reveal molecular differences between COPD subtypes.

We identified 124 COPD subtype-associated proteomic biomarkers in COPDGene and validated these associations using several approaches. Firstly, ANOVA and post-hoc SNK tests including participants classified as COPD subtypes and control supported most of the same proteomic biomarkers. Secondly, 64 of these 124 COPD subtype-associated proteins were nominally validated with the same effect direction in SPIROMICS. We compared the directionality of associations among the 124 EPD vs. NEPD COPD protein biomarkers in COPDGene and SPIROMICS, and 120 had the same direction of effect. Thus, we found high concordance between COPDGene and SPIROMICS, supporting the validity of these proteomic biomarkers for COPD subtypes. The identification of several significant biological pathways related to EPD vs. NEPD COPD protein biomarkers, and a correlation network module associated with EPD vs. NEPD suggest that these EPD vs. NEPD COPD-associated proteins are involved in shared biological processes related to COPD heterogeneity. Interestingly, most of the EPD vs. NEPD COPD protein biomarkers had higher plasma levels in participants with NEPD; the aetiology of this phenomenon is unknown but could relate to distinct pathobiological processes in EPD and NEPD.

Among five proteins with COPD subtype association FDR < 0.05 in both COPDGene and SPIROMICS, the soluble RAGE (sRAGE) protein biomarker (encoded by the AGER gene) acts as a decoy to block inflammation.^49,50 In 2011, Miniati reported lower sRAGE levels in patients with COPD.⁵⁰ In 2021, Keefe et al. utilised an integrative genomic approach to identify sRAGE as a causal and protective biomarker for lung function.⁵¹ Pratte et al. associated sRAGE with airflow obstruction and emphysema.⁵² Lower circulating sRAGE is related to severe airflow obstruction and increased emphysema in COPD,^51,52 consistent with our results. CD93 is a myeloid cell biomarker associated with intercellular adhesion and inflammation^53,54 by activating beta-1 integrin through the CD93/Multierin-2/active beta-1 integrin complex.⁵⁵ Plosa and Zent confirmed that beta-1 integrin regulates lung inflammation, and beta-1 integrin deletion in type 2 alveolar epithelial cells leads to the development of emphysema in mice.⁵⁶ Another biomarker, KLK10, has been reported to be associated with FEV1% predicted based on blood methylation analysis in COPD.⁵⁷ As for other proteomic biomarkers with nominal replication in SPIROMICS, PIGR (Polymeric immunoglobulin receptor) binds to polymeric IgA.⁵⁸ Gohy et al. recognised that PIGR is downregulated in COPD, and the downregulation is associated with COPD severity.⁴³ An airway innate immunity study presented by Richmond et al. utilised mouse models to investigate the underlying mechanisms for persistent airway inflammation in COPD.⁵⁹ PIGR deficiency was shown to be associated with innate immune activation, which further drives progressive small airway remodelling and emphysema. Other nominally replicated associations for EPD vs. NEPD COPD included terminal pro-BNP; these findings may reflect key differences in comorbidities between participants with EPD and NEPD COPD, including heart disease.⁹

Gene set and gene ontology enrichment analyses revealed potential biological functions for significant COPD subtype proteomic associations, including epithelial mesenchymal transition (EMT), cell adhesion, and collagen-containing extracellular matrix. Su et al. noted that EMT is an important pathophysiological process involved in airway fibrosis, airway remodelling, and malignant transformation of COPD associated with cigarette smoking.⁶⁰ Fujioka et al. identified the potential of human adipose-derived mesenchymal stem cells to ameliorate elastase-induced emphysema through mesenchymal-epithelial transition in mouse models.⁶¹ For cell adhesion, Mimae et al. reported that a cell adhesion molecule in lung tissue, CADM1 (Cell Adhesion Molecule 1), activated alveolar cell apoptosis in emphysematous lungs,⁶² and the apoptosis of alveolar cells can further contribute to emphysema development.⁶³ Karakioulaki et al. reviewed the role of extracellular matrix remodelling in COPD and confirmed that collagen remodelling plays an important role in emphysema development,⁶⁴ which was experimentally verified in mouse models decades ago.⁶⁵

We identified thirteen COPD subtype protein biomarkers representative of a correlation network module that was associated with COPD subtypes. Highly connected module components for this COPD subtype-associated module include TNFRSF1A, EFNA2, and EFNB2. Fujita et al. recognised the critical role of TNFRSF1A in the pathogenesis of pulmonary emphysema in mouse models.⁶⁶ Soluble ephrin-B2 was identified as a profibrotic mediator in lung fibrosis.⁶⁷ EFNA2 has been reported to be associated with weight loss in COPD, and cachexia is a key systemic effect in some advanced patients with COPD.⁶⁸ Thus, correlation network analysis revealed molecular differences between EPD vs. NEPD COPD from multiple functional perspectives.

Genetic associations for COPD subtypes and COPD subtype-associated proteomic biomarkers were also evaluated. Due to the modest sample size for a genetic association study, only previously identified COPD genetic determinants were included. A SNP in LRMDA was associated with EPD vs. NEPD COPD, and six other SNPs had nominal EPD vs. NEPD COPD associations (including the Z allele of SERPINA1). Thus, some COPD genetic determinants appear to have differential impact in EPD and NEPD COPD. In addition to the previously reported genetic control of sRAGE,^20,21,27 we found that sphingolipid associated protein (GM2A) levels were associated with a local COPD GWAS SNP (rs979453) in control and NEPD participants but not in EPD COPD, suggesting that GM2A is a functional COPD susceptibility gene. Key genes in the chromosome 5q region have not been reported, but our analyses indicated that GM2A may be at least one of the functional genes in this region. Kechris and Bowler reported significant associations between sphingolipids and emphysema in plasma,⁶⁹ which potentially could provide a connection between GM2A and emphysema.

Clinical characteristics including treatment with corticosteroid medications and comorbidities such as cardiovascular disease may also affect COPD proteomic associations. Corticosteroids have strong immune-modulating effects.⁷⁰ We observed that corticosteroids are more frequently used in patients with emphysema-predominant COPD, which may be due to the relatively lower lung function of participants with emphysema-predominant COPD. Cardiovascular diseases frequently coexist with COPD.⁷¹ We found that five previously reported cardiac biomarkers, including NPPB, DSC2, CST3, KRT1, and ADH1B, were differentially expressed in COPD subtypes, with higher levels in participants with NEPD. This is consistent with the increased risk for cardiovascular events in participants with NEPD in relation to elevated coronary artery calcification scores that we recently reported in COPDGene.⁷²

Although we have performed a comprehensive proteomic analysis of two COPD subtypes, there are some limitations in our current study. One potential limitation is that we focused on plasma rather than lung proteomics; plasma proteomic biomarkers may not reflect the molecular heterogeneity within the lungs, and the cells and tissues producing these plasma proteins are uncertain. The use of plasma proteomics can capture systemic differences between COPD subtypes but may also introduce bias in pathway analyses. Plasma proteomics was evaluated using SomaLogic SomaScan® panel; however, SomaScan is an aptamer-based proteomic platform with potential off-target cross-reactivity and limited capacity for isoform detection. Since the panel includes a minority of human proteins, additional COPD subtype biomarkers are likely to be discovered with more comprehensive proteomic assessments. For disease subtyping, the EPD/NEPD classification does not capture all aspects of COPD heterogeneity (e.g., emphysema pathological pattern and distribution). The attenuation of some of the identified protein biomarker associations after adjusting for FEV1 (% predicted) suggests that some of these biomarkers may relate to disease severity. Another limitation is related to unmeasured cofounding factors for evaluating proteomic disease associations. In our analyses, we included several known SomaScan methodological covariates (clinical centre and cellular components) and widely used COPD clinical epidemiological covariates (age, sex, self-reported race, pack-years of smoking, current smoking status, and BMI) as confounders. However, there may still be unknown methodological effects (e.g., unknown technical factors for SomaScan analysis) and biological effects (clinical/biological covariates for emphysema pathogenesis). Although individuals with recent COPD exacerbations were not included, recent plasma and systemic infections could also influence our proteomic results. Inaccurate reports for self-reported covariates like smoking status may further affect our results. In our study, we tried to overcome these limitations by replicating proteomic associations in an independent cohort, SPIROMICS. Using an independent cohort not only validates our findings on COPD subtype proteomic associations but may also limit the effects of unmeasured confounders.

In plasma proteomics analyses, multiple biomarkers associated with emphysema-predominant vs. non-emphysema predominant COPD subtypes were identified. These COPD subtype proteomic biomarkers were enriched for biological functions including epithelial mesenchymal transition, cell adhesion, and collagen-containing extracellular matrix. Correlation-based network analyses identified a network module significantly associated with EPD vs. NEPD COPD. We also observed significant genetic effects on COPD subtypes and their associated proteomic biomarkers. Further study will be required to determine whether any of these identified proteomic biomarkers trigger inciting events in COPD pathogenesis that lead to disease heterogeneity.

Contributors

Conceptualisation: Edwin K. Silverman, Yu-Hang Zhang, Peter J. Castaldi, Russell P. Bowler, Katherine A. Pratte, Gregory L. Kinney, Kendra A. Young, Sharon M. Lutz, Craig P. Hersh, Michael H. Cho, Jarrett D. Morrow, Heena Rijhwani.

Data curation: Edwin K. Silverman, Yu-Hang Zhang, Peter J. Castaldi, Russell P. Bowler, Katherine A. Pratte, Michael H. Cho, Jarrett D. Morrow, Edwin K. Silverman.

Formal analysis: Edwin K. Silverman, Yu-Hang Zhang.

Funding acquisition: Edwin K. Silverman.

Investigation: Edwin K. Silverman, Yu-Hang Zhang.

Methodology: Edwin K. Silverman, Yu-Hang Zhang, Peter J. Castaldi, Katherine A. Pratte, Sharon M. Lutz, Michael H. Cho, Jarrett D. Morrow.

Project administration: Edwin K. Silverman.

Resources: Edwin K. Silverman, Russell P. Bowler, Katherine A. Pratte.

Software: Edwin K. Silverman, Yu-Hang Zhang, Katherine A. Pratte.

Supervision: Edwin K. Silverman.

Validation: Edwin K. Silverman, Russell P. Bowler, Yu-Hang Zhang.

Visualisation: Edwin K. Silverman, Russell P. Bowler, Yu-Hang Zhang.

Writing: Edwin K. Silverman, Yu-Hang Zhang, Peter J. Castaldi, Russell P. Bowler, Katherine A. Pratte, Gregory L. Kinney, Kendra A. Young, Heena Rijhwani, Sharon M. Lutz, Craig P. Hersh, Michael H. Cho, Jarrett D. Morrow.

Authors have directly accessed and verified data reported in the manuscript: Edwin K. Silverman, Yu-Hang Zhang, Peter J. Castaldi, Russell P. Bowler, Katherine A. Pratte, Gregory L. Kinney, Kendra A. Young, Sharon M. Lutz, Craig P. Hersh, Michael H. Cho, Jarrett D. Morrow.

All authors contributed substantially to the study design, data analysis, or interpretation. All authors reviewed the manuscript for intellectual content, contributed to revisions, and approved the final version for submission.

Data sharing statement

The COPDGene and SPIROMICS deidentified participant data sets and data dictionaries are available for monitored public access through dbGaP (Database of Genotypes and Phenotypes). More information about the study and how to access SPIROMICS data is available at www.spiromics.org.

Declaration of interests

The authors report the following conflicts of interest:

In the past three years, EKS has received grant support from Bayer and Northpond Laboratories. CPH has received grant support from Alpha-1 Foundation, Bayer, Boehringer-Ingelheim, and Vertex, and consulting fees from Chiesi, Ono, Sanofi, and Takeda. JDM has received support from an Alpha-1 Foundation Research Grant. MHC has received grant support from Bayer. PJC has received grant support from Sanofi and Bayer, and consulting fees from Verona.