Background

Chronic Obstructive Pulmonary Disease (COPD) is defined by the Global Initiative for Chronic Obstructive Lung Disease (GOLD) as airflow limitation that is not fully reversible[1]. This deliberately broad and simple definition based on reduced expiratory airflow has been useful in leading to increased awareness and understanding of the disease[2]. However, substantial heterogeneity within this definition exists[3, 4]. Moving beyond spirometry and evaluating other variables is critical to understanding differences in patients with COPD, to gain mechanistic insights into the disease, to identify those at highest risk of specific outcomes, and to personalize therapy [4–7].

Substantial evidence indicates that genetic variation contributes to differences in COPD susceptibility; however, replication of genetic associations in COPD - and in many other complex diseases - has generally been poor[8]. Disease heterogeneity is likely an important factor for these inconsistent findings[9, 10]. Several attempts to overcome heterogeneity have been used, including using classic subtypes of chronic bronchitis or emphysema[11], defining subtypes based on a pathophysiologic characteristic (such as rapid or slow decline in lung function[12]), or assessing phenotypic characteristics by chest CT scans[13].

Statistical learning techniques offer an opportunity to extract novel patterns and trends from phenotypic data[14], and thus identify COPD subtypes without using a priori expectations about disease characteristics[15–20]. To our knowledge, these strategies have not been applied in a group with severe emphysema, nor have studies used these subtypes in a genetic association study. We hypothesized that cluster analysis would identify distinct subtypes of COPD subjects, and that variants in COPD candidate genes would be associated with these subtypes.

Methods

Details of subject recruitment and phenotyping in the National Emphysema Treatment Trial (NETT) have been reported previously[21]. Briefly, NETT participants had physician-diagnosed COPD, FEV1 ≤ 45% predicted, evidence of hyperinflation on pulmonary function testing, and bilateral emphysema on chest CT scan. Enrollment in the NETT Genetics Ancillary Study began after the initiation of the clinical trial, and thus only a subset of the original cohort had DNA available for genotyping. The characteristics of NETT subjects included and excluded from this analysis are shown in Additional File 1, Table S1. Participants gave written informed consent. The appropriate institutional review boards approved all studies. Self-identified white subjects in the NETT Genetics Ancillary Study with complete CT phenotypic data (emphysema and airway wall quantitative measures) were included in the analysis.

We selected a set of 31 CT, lung function, and other key phenotypic variables, based on clinical relevance, inclusion in previous genetic association studies, and complete data, to avoid subject drop-out (Table 1). Measurements of the phenotypic variables have been previously described[21–25]. 12 SNPs from 5 genes were chosen on the basis of available genotyping and prior associations with COPD (Table 2), and included genes involved in xenobiotic metabolism (EPHX1 and GSTP1) and surfactant homeostasis (SFTPB), as well as two genes identified in part through linkage studies: TGFB1, a cytokine growth factor, and SERPINE2, a thrombin and urokinase inhibitor. A limited number of SNPs were selected in order to limit multiple statistical testing in this relatively small study population[26]. For the gene SERPINE2[27, 28], in which numerous associations have been described, we chose SNPs tagging associations found in at least two populations, using a r2 cutoff of > 0.8 in Tagger [29] as implemented in Haploview 4.1 [30].

Table 1 Factor analysis.
Full size table
Table 2 Single nucleotide polymorphisms (SNPs).
Full size table

We used factor analysis as a guide to determine which COPD phenotypic variables to include in our clustering analysis[31]. Factor analysis is a data reduction technique related to principal component analysis, where shared variability in several observed variables is explained in terms of fewer unobserved variables, called factors. The strength of the relationship between the observed variables and factors can be measured by factor loadings. We used factor analysis in two ways: first, to select variables which represent greater amounts of shared variability; and second, among these variables, to select one representative measurement using a high factor loading to avoid over-weighting correlated COPD characteristics, which could bias a cluster analysis.

The goal of cluster analysis is to assign subjects to groups, where subjects in the same cluster are more similar to each other than they are to subjects in other groups[14]. Similarity is generally defined using a measurement of distance, calculated using the difference between measurements. As numerous clustering methods exist, we evaluated the performance of several clustering algorithms. We chose the best performing clustering technique and cluster number using the silhouette width, a measure of how close each point in one cluster is to points in neighboring clusters. We then examined each variable for differences among the clusters.

Results

A total of 308 subjects from the NETT Genetics Ancillary Study were included in the analysis. Phenotypic variables selected for inclusion in the factor analysis are shown in Table 1. SNPs included for analysis, along with their minor allele frequencies and previous genetic association studies in COPD (both in the NETT cohort and others) are listed in Table 2. Six factors accounted for 75% of the common variance. Eigenvalues for these factors ranged from 1.7 to 4.9; two additional factors (not shown) had eigenvalues > 1.0. The results of the factor analysis are shown in Table 1. These factors were interpreted as: 1) spirometry (containing pre- and post-bronchodilator FEV1 and FVC percent predicted); 2) airway wall thickness (wall thickness, derived square root wall area of a 10 mm internal perimeter airway and derived wall area percent of a 10 mm airway); 3) FEV1/FVC ratio; 4) quantitative emphysema severity and distribution (divided into equal thirds by absolute lung height from apex to base), using a cutoff of -950 Hounsfield units; 5) bronchodilator responsiveness; and 6) maximum work and gender. The following variables were chosen as representative based on relatively high factor loadings, accounting for a greater proportion of shared variability: post-bronchodilator FEV1 percent predicted (for factors 1 and 3), airway wall thickness (factor 2), apical emphysema (factor 4), and bronchodilator responsiveness (factor 5).

Based on measures of silhouette width, the k-means clustering algorithm using four clusters was found to be optimal for this dataset. This optimal value was low (~0.2) consistent with modest separation of the clusters. A plot of the clusters with each pair of phenotypic variables is shown in Additional file 1, Figure S1. Differences between the selected phenotypic variables by cluster are shown in Table 3. As expected, all selected phenotypic variables selected for use in cluster analysis were significantly different between clusters (P < 10-3). Cluster 1 had the greatest degree of emphysema and the least airway wall thickness, as well as a lower FEV1. Conversely, cluster 4 had the highest airway wall thickness and the least emphysema, and also had lower bronchodilator responsiveness and FEV1. Cluster 2 was a milder subgroup, with the highest FEV1 and bronchodilator responsiveness, as well as less emphysema. Cluster 3 also had less emphysema and in addition, less airway wall thickness; however, in contrast to cluster 2, this cluster was more severely affected with the lowest FEV1 and bronchodilator responsiveness.

Table 3 Baseline Phenotypic Characteristics and Results of Cluster Analysis.
Full size table

Additional phenotypic variables, not included in the cluster analysis, were then examined to determine other characteristics of these clusters. Significant differences among the groups were found for several characteristics (Table 3). Cluster 1, the emphysema predominant cluster, had a lower BMI, fewer pack-years of smoking, higher total lung capacity, and lower diffusing capacity, along with a lower six minute walk distance and maximum work. Consistent with the radiographic clustering and the factor loadings, CT emphysema severity and apical-basal emphysema difference (defined as absolute percent emphysema in the upper lung region minus the lower lung region) were more severe, while airway wall area and square root of wall area were lower. Conversely, cluster 4, the airway predominant cluster, had a higher BMI, lower total lung capacity, and less severe emphysema and higher airway wall measurements, with a lower PaO2 and lower six minute walk distance. Cluster 2, the milder severity, bronchodilator responsive subtype, had a higher BMI, greater FVC and DLCO, a lower PaCO2, higher six minute walk distance and maximum work, fewer symptoms of dyspnea, and fewer exacerbations, despite being of slightly older age. Cluster 3, with a lower FEV1 despite less severe radiographic emphysema and airway wall thickness than the other clusters, had more dyspnea and a higher PaCO2, with slightly younger age.

To determine whether specific SNPs were associated with cluster membership, we tested genotypes for each of the 12 candidate gene SNPs with cluster membership. A chi-squared P value of 0.034 was seen for a SNP in TGFB1, rs1800470; no other P values were nominally (<0.05) significant. In pairwise testing using an additive model of each cluster versus all other clusters, this SNP was associated with membership in cluster 1 (P = 0.002).

Further details on study methods and additional results are available in Additional File 1, including plots of correlations and cluster separation in two-dimensional space (Additional file 1, Figures S1 and S3).

Discussion

Despite the description of COPD subtypes more than 40 years ago[32] and substantial progress since then in understanding COPD-related phenotypes[33, 34], only a few attempts have been made to use statistical methods to define novel COPD subtypes[15, 16]. Using a large, well-characterized set of subjects with severe emphysema, we demonstrate the potential utility of using statistical learning methods to find relationships among phenotypic and genotypic characteristics to elucidate disease heterogeneity.

Several methods have attempted to address issues of disease heterogeneity in obstructive airway diseases. Statistical learning techniques such as factor analysis have been used to reveal novel insights into characteristics such as dyspnea or inflammation in COPD[20, 35–37]. Cluster analysis has confirmed classic chronic bronchitis and emphysema subtypes[15] or illustrated overlap of characteristics of COPD and asthma[16], and a combination of factor analysis and cluster analysis has defined asthma subtypes[31]. These techniques show promise in identifying disease subtypes (subsets of subjects), or intermediate disease-related phenotypic characteristics (endotypes/endophenotypes[38]). Endophenotypes have already been of substantial utility in genetic association studies in psychiatry[39].

To date, however, there has been limited use of disease subtypes in genetic association studies in COPD. Investigators have tested for specific associations with classic subtypes[11, 40, 41], or with specific disease-related phenotypic characteristics such as emphysema distribution[22] or functional measures[42]. Factor analysis has been used to demonstrate differences in heritability of components of asthma[43]. Cluster analysis is frequently used in gene expression, and such analyses have been used to define subtypes - though these subtypes have not always been clearly associated with the available clinical characteristics[44]. Our study demonstrates the potential utility of statistical learning methods in the heterogeneous syndrome of COPD.

Our cluster analyses identified four subtypes of subjects in this cohort with severe emphysema: 1) emphysema predominant, 2) milder severity, bronchodilator-responsive, 3) discordant lung function/CT emphysema and airway severity, and 4) airway predominant. Some of the phenotypic associations in these groups, such as a lower BMI with more severe quantitative CT emphysema, have been previously seen[13, 45], while others, such as a higher bronchodilator responsiveness in the group with higher FEV1, differ from previous reports[46, 47]. The association of the nonsynonymous Leu10Pro TGFB1 SNP rs1800470 with cluster 1 is consistent with a previously reported association of apical emphysema in this cohort [22] and association of this SNP with reduced lung function has also been seen in a Japanese emphysema cohort[48]. Notably, this SNP has been demonstrated to be of functional significance, with the G allele (C on the reverse strand) resulting in increased production of TGFB1[49]. Several studies have demonstrated an increase in TGFB1 both in the lung[50–52] and in plasma[53] in subjects with COPD, as well as a relationship between TGFB1 levels and lung function, though the relationship between these findings and the rs1800470 genotype is not entirely clear[53].

Conversely, most of the previously reported SNP associations with COPD-related phenotypic characteristics did not demonstrate associations with our clusters. Nonsignificant findings could be due to loss of power from categorical cluster assignment and resulting small sample size, and the use of an omnibus test for genetic association. More importantly, our analysis attempts to determine whether genetic variants lead to a subtype of COPD subjects which share a set of phenotypic characteristics; as such, it does not attempt to determine the specific genotypic-phenotypic variables whose relationship leads to a significant association. Whether one of these approaches - association analysis with individual phenotypic characteristics, or with subtypes of subjects- is superior in identifying replicated genetic associations, or whether the approaches are separately informative, remains to be seen.

Our study has several strengths. First, we used relatively unbiased methods, in both factor analysis and cluster analysis, to select uncorrelated variables and determine severe COPD subtypes using the rich set of phenotypic and quantitative measures available in NETT. Second, our analysis is the largest reported cluster analysis using CT phenotypic variables. Third, despite our homogeneous study population, we were able to discern emphysema subtypes, which differed on variables not used to perform clustering. While all four of these subtypes have not previously been identified, our emphysema and airway-predominant clusters are consistent with a priori defined subtypes used in previous studies[13]. Importantly, recent evidence shows that airway wall thickening and emphysema aggregate independently in families of individuals with COPD[54], suggesting that recognizing these differences may be important for discovering genetic associations.

Our results should be regarded as exploratory for several reasons. First, our dataset was based on available NETT data. Specific relationships between variables - for example, the high correlation between apical and total emphysema - may be due to selection biases of the NETT population. NETT subjects were likely biased towards those without predominant airway disease, and CT scans were suboptimal for assessment of airway wall remodeling due to the thicker slices associated with pre-MDCT (multi-detector CT) imaging. Similarly, our genotypic data was limited to a pre-specified subset of previous positive associations in candidate genes, and our cohort was limited to those enrolled in the NETT Genetics Ancillary Study (Additional File 1, Table S1). Our selection of phenotypic and genotypic variables for inclusion was strongly influenced by the limitations of available data, and decisions were made based on clinical judgement of relevance.

Second, our analysis also found that the separation of clusters was weak, indicating segmentation and not a true separation of these subtypes using clustering. Correspondingly, we found no strong evidence of smaller groups of more distinct subtypes. Furthermore, the small size of our clusters limits the power of association analysis, and our association with rs1800470 was not corrected for multiple comparisons. Given these limitations in this relatively homogeneous cohort, an attempt to validate these findings of specific subtypes using these or similar methods in other well-phenotyped COPD cohorts should be performed. Using a more heterogeneous and less selected group of subjects, in combination with improved radiographic measures, may result in more pronounced and distinct subpopulations.

Conclusions

The volume of genetic and phenotypic information available in COPD cohorts is rapidly increasing; the number of potential relationships between phenotypic and genotypic characteristics increases exponentially. Statistical learning techniques using multivariate methods, such as dimension reduction and cluster analysis, have the potential to assist in analyses of these complicated problems. Our study demonstrates that application of these techniques, even in a highly selected group of subjects with severe emphysema, has the potential to elucidate phenotypic heterogeneity and disease pathophysiology.