Distinct COPD subtypes in former smokers revealed by gene network perturbation analysis

The COPDGene Study is a longitudinal study that aims to investigate the genetic basis of COPD susceptibility and progression. Our subtype discovery dataset consisted of 1211 COPDGene subjects for whom whole blood RNA-seq data were collected at the 5 year follow-up visit [25]. The first validation dataset included 1444 COPDGene participants that were sequenced later. These samples were not included in the training dataset and were processed independently. The second validation dataset consisted of 821 unrelated MESA participants. MESA is an ongoing prospective cohort study that recruited over 6000 participants in six communities throughout the United States between 2000 and 2002 [26]. Peripheral blood mononuclear cell (PBMC) gene expression was measured by RNA-seq at Exam 5 between 2010 and 2012. Detailed phenotype data (including spirometry and CT scan) were also collected at this exam.

The COPDGene RNA-seq samples were preprocessed as in the Additional file 1: Methods. The reference gene network was built on a group of former smokers, selected conservatively based on the following criteria (Fig. 1A): participant had (a) both Phase 1 (baseline) and Phase 2 (5-yr) visits; (b) normal spirometry both visits; (c) less than 5% percent emphysema in both visits (LAA-950); (d) less than 5% decrease in percent predicted FEV₁ between the two visits. This filtering resulted in 128 reference samples (training dataset). To increase power, all remaining samples from participants who formerly smoked cigarettes and who did not meet the criteria for the reference group were included in the disease group, leaving 489 samples for subtype discovery.

For the COPDGene validation dataset, the same filtering criteria identified 149 control and 614 COPD former smoker samples. The MESA reference group was selected with similar criteria, except the threshold for FEV₁ decline between Exam 3/4 and Exam 5 (< 3% predicted). Participants with no spirometry data were included in the non-reference group. This resulted in 104 reference and 327 case MESA samples. Table 1 summarizes the characteristics of these three study groups.

We used ssNPA, a network-based disease subtyping method, to learn COPD subtypes in the discovery cohort (Fig. 1B). The details of the process are presented in Additional file 1: Methods. To investigate the clinically relevant differences among the clusters of COPD samples we identified, we compared the values of 105 clinical variables across the clusters (Additional file 2: Table S1). These included spirometry, chest CT scan, symptom questionnaires, and white blood cell differentials as well as medical history, medication, and comorbidity information. The p-values were calculated using Kruskal–Wallis (continuous, ordinal) and Chi-squared test (binary). A false discovery rate (FDR) threshold of q-value < 0.05 was used for multiple test correction. Pairwise comparisons between cluster means were assessed by Wilcoxon test. Survival analysis was performed by the Kaplan–Meier method with the survfit function from the survival R package (v. 3.1-8).

To better understand how the clusters were separated, we considered the magnitude of the PCA loading for each feature. Gene features with the highest loading values in the top principal components correspond to the genes whose deregulation relative to the controls contributes the most to separating the clusters.

Finally, MESA and COPDGene validation cohort samples were assigned to one of the four subtypes using the following procedure described in Additional file 1: Methods.

ssNPA separated the 489 COPD PBMC samples (all former smokers) into four clusters (Fig. 1C, Additional file 1: Fig. S8). The first two clusters were of roughly equal size (cluster 0 and cluster 1 have 37.0% and 31.9% of the samples, respectively). Cluster 2 and cluster 3 were smaller (with 16.2% and 14.9% of the samples). Mean age was similar in all clusters. Cluster 2 had a higher proportion of women and its subjects had a higher neutrophil and lower eosinophil percent, which are indicative of inflammation.

Statistical analyses identified clinical characteristics that are significantly associated with these clusters (Table 2). Overall, cluster 2 represented the most impaired subjects, with lower 6-min walk distance, more symptoms (measured by COPD Assessment Test, CAT) and dyspnea (measured by the Modified Medical Research Council Dyspnea Scale, MMRC), and the worst disease-related quality of life on the St. George’s Respiratory Questionnaire (SGRQ). However, spirometry (FEV₁ percent predicted and FEV₁/FVC) was similar in clusters 2 and 3, but worse than cluster 0 and 1. Clusters 0 and 1 have the best spirometric lung function, but ~ 4 times more subjects in cluster 0 are using inhaled corticosteroids (5.08% vs 1.3%; Table 2). Regardless, the percentage of corticosteroid usage in subjects of cluster 2 and cluster 3 is even higher (11.54%, 13.7%), making this feature significantly different between clusters. Diffusing capacity of carbon monoxide (DLCO) percent predicted was highest in cluster 0 (76.7%), but similar in the other 3 clusters (70.8, 69.1, 70.5% predicted, respectively). Cluster 3 had the greatest percent emphysema (12.03%), followed by cluster 2 (10.84%), while cluster 0 and cluster 1 had the least emphysema (8.59%, 8.47%) (Additional file 2: Table S1). However, these differences are not significant at the 5% FDR level (q-value = 0.068). Similarly, observed differences in Pi10 and Perc15 are not significant (q-value = 0.072–0.076).

Survival analysis shows significant differences across the four clusters of COPDGene participants (Fig. 2; p-value < 0.001). This significance was driven by cluster 2, which has the worst outcome with the steepest decline in survival probability. Despite a lower FEV1 in cluster 3, survival probabilities in this cluster and cluster 1 began to decline similarly after approximately 2 years following study visit 2 to an intermediate level. Cluster 0 maintained the highest survival probability after this time. The observed survival differences between clusters 0, 1, and 3 are not significant. We also note that clustering based on spirometry characteristics alone (e.g., FEV1 percent predicted) did not produce differences in mortality (Additional file 1: Fig. S9).

We are aware that comorbidities may drive mortality either directly or through medications that people take because of them. Although our sample subtyping was based on blood gene expression and any comorbidities or medications that may influence our subtyping had to exert their effect through gene expression, we tested to see if our subtypes differ in terms of comorbidities. We found that none of the comorbidities recorded in COPDGene was significantly different between clusters (Additional file 1: Table S3.)

We further look in the variables that significantly differ between cluster 0 and cluster 1 (the two clusters with the highest FEV₁). Both DLCO and the change in 6-min walk distance between the two visits are significantly worse in cluster 1 compared to cluster 0 (p-value = 0.011 and 0.017, respectively). DLCO differences are interesting given that percent emphysema was similar in the two clusters (Additional file 3: Table S2), which may indicate another process such as pulmonary vascular disease. A marker of pulmonary hyperinflation (FRC/TLC ratio) and a fatigue symptom also were found to be significantly different between these clusters (p-value = 0.005 and 0.036, respectively). However, the correlation between FRC/TLC ratio and adjusted DLCO percent predicted in each of these two clusters was quite low (− 0.41 and − 0.53 for cluster-0 and cluster-1, respectively), which is in line with previous studies [27, 28]. One explanation is that DLCO captures more than just emphysema, including pulmonary vascular disease and small airway disease [29].

We validated the ssNPA-generated subtypes by scoring new COPD samples from the validation groups using the model we learned from the COPDGene training cohort. We chose this type of validation because it demonstrates the ability to correctly assign new samples to these four subtypes, which is an attractive feature for future clinical application of these subtypes. Out of remaining 614 COPD samples in the COPDGene validation set, 288 were assigned to cluster 0, 193 to cluster 1, 42 to cluster 2, and 91 to cluster 3 (Additional file 1: Fig. S10). We emphasize that this cluster assignment was based solely on gene expression data. Therefore, we next checked to see if the clinical features that differed across clusters in the primary analysis followed the same trends across the clusters of these new samples. In general, we observed the same trends for the significant features when we looked at the discovery set of samples compared to the validation set of samples (Fig. 3), and many pairwise relationships between the cluster means of these features were replicated in the validation set (Additional file 1: Fig. S11). In general, we see higher concordance between the two cohorts when it comes to symptoms (CAT, MMRC Dyspnea and SGRQ Total scores) as well as forced expiratory flow (FEF 25–75).

The validation in the MESA cohort produced a similar result. 150 samples were assigned to cluster 0, 64 were assigned to cluster 1, 41 to cluster 2, and 72 to cluster 3 (Additional file 1: Fig. S12). Cluster 2 contained the most severe COPD phenotypes and contained the fewest MESA samples, and, conversely, cluster 0 exhibited the mildest disease and contains the most MESA samples. Again, we observed the same general trends for the clinical features in the MESA validation analysis as we did in the COPDGene discovery analysis (Fig. 3). Fewer of the pairwise comparisons of the means from the cluster of the clinical features were replicated in MESA, likely due to the smaller number of participants for whom these phenotypes were measured (Additional file 1: Fig. S13).

We further analyzed the identified COPD subtypes to investigate the molecular differences among them, which lead to different clinical phenotypes. We examined the gene deregulation features that had the largest PCA loadings to identify the genes that make the largest contribution to the clustering of the COPDGene discovery dataset (Additional file 1: Table S4). We focused on the top five loadings for each of the first six principal components (PCs) that were used to cluster the subjects and found that many of the genes came up more than once, including DSP and GSTM1, which have been previously associated with COPD. For many of these genes, we observe large differences across clusters in the distributions of deregulation magnitude (Additional file 1: Fig. S14), even between clusters whose subjects have similar lung function.