Network medicine analysis of COPD multimorbidities

Clinical experts selected 16 prevalent diseases frequently associated with COPD (Table 1) [12]-[14]. Their Concept Unique Identifiers (CUI) were obtained from the Unified Medical Language System (UMLS) Metathesaurus (Figure 1 panel A) [15]. After manual curation, the disease vocabulary used in the analysis included a mean of 59 CUIs (range 4–332) for each disease (and 14 CUIs for COPD, 33 for emphysema and 16 for chronic bronchitis (Additional file 1: Table S1)).

The “diseasome” is a network representation of the associations between diseases, where two of them are linked if they share alterations in genes or proteins, or if the altered proteins are connected in the interactome (9) (Figure 2). DisGeNET (version 2.0), a database that integrates information on human diseases and their associated genes from various public databases and from the biomedical literature [16], was used to collect information on proteins altered in COPD and the multimorbidities studied here using our disease vocabulary (Figure 1 panel B). The human interactome was obtained by combining PPI information from the Pathway Commons database [17] (http://​www.​pathwaycommons.​org/​), gathering biological pathway information collected from public pathway databases and HIPPIE [18] (http://​cbdm.​mdc-berlin.​de/​tools/​hippie/​information.​php), a human PPI database that integrates multiple experimental PPI datasets (Figure 1 panel B).

When building the diseasome based on proteins in common between two diseases, it could be argued that it is more likely to find shared proteins among diseases that have been characterized more extensively. In order to minimize this bias when estimating the strength of the association between two given diseases (dis1 and dis2) in the diseasome (Figure 1 panel C), we calculated the Molecular Comorbidity Index (MCI) as defined by:

where proteins _dis1 and proteins _dis2 are the proteins associated with disease 1 and 2, respectively, proteins _dis1→dis2 are those proteins associated with disease 1 that interact with those associated with disease 2 (and vice versa (proteins _dis2→dis1 )), ∩ is the intersection operator and ∪ is the union operator between two sets of elements (proteins _dis1 and proteins _dis2 ). The sets resulting in both numerator and denominator are written within vertical bars to indicate their cardinality (number of element).

To identify the most significant biological functions of the shared proteins between two diseases, a functional enrichment analysis with biological pathways from Reactome was performed [19]. The Reactome database contains information on genes, proteins and their participation in biological pathways. In our analysis, we used the R package ReactomePA, which uses the hypergeometric function to test the significance of annotations [20]. Significant annotations were those with q-values ≤ 0.05 (the q-value corresponds to the false discovery rate (FDR), an adjusted equivalent to the standard statistical p-value incorporated in ReactomePA).

To evaluate the number of biological pathways shared by COPD multimorbidities, we used the Jaccard coefficient, which is defined as:

where multimorbidity ₁ and multimorbidity ₂ represent two pairs of COPD multimorbidities (for instance lung cancer-COPD and diabetes-COPD) whereas pathways of multimorbidity ₁ and pathways of multimorbidity ₂ represent the biological pathways in which the proteins associated with the pairs multimorbidity ₁ and multimorbidity ₂ participate respectively. The Jaccard Coefficient is a measure of the degree of similarity between two COPD multimorbidities (for instance lung cancer-COPD and diabetes-COPD) at the level of biological pathways. Results were visualized as heat-maps using Gitools [21].

Using the Comparative Toxicogenomics Database (CTD), a database gathering information about gene/protein and chemical interactions, we investigated if the genes and proteins shared by COPD multimorbidities were potential biological targets for chemical compounds present in the tobacco smoke [22],[23].

COPD has been associated in the literature with alterations in 187 genes, whereas the number of genes associated with the 16 comorbidities analyzed here range from 19 (cachexia) to 1,369 (lung cancer) (Table 1 and Additional file 2: Figure S1 in the online data supplement). As explained in Figure 2, we used this knowledge on the genetic basis of these diseases in combination with data obtained from the interactome to build the COPD diseasome, and found that all 16 comorbidities studied here can be connected through shared genes and/or proteins that interact within cellular networks. Figure 3 illustrates how COPD and anemia are linked through this network of interactions. Both diseases are directly associated to 3 proteins in common, SOD2, TP53 and TNF (green nodes in Figure 3), and are also connected through the interactions between 22 proteins (pink and lila nodes in Figure 3). These bridging proteins constitute potential candidates to explain the association between COPD and anemia that should be further investigated.

In the COPD diseasome, the number of shared genes or proteins ranged from less than 100 (COPD and anemia) to more than 700 (COPD and lung cancer) (Figure 4). In terms of the Molecular Comorbidity Index (MCI), that is an indicator of the strength of the association between two diseases normalized for those better studied (for instance, lung cancer, see Methods), ischemic heart disease ranked first in the list of COPD multimorbidities (approximately 400 shared proteins with a MCI of 0.53), followed by lung cancer, stroke and diabetes. The frequency distribution of the MCI (Figure 4) paralleled that reported in the literature for the multimorbidities studied here [10],[24],[25].

The proposed network medicine workflow not only allows a global analysis of the genes and proteins supporting a particular multimorbidity (see next sections), but it also provides detailed information on specific genes and their association to the multimorbid diseases under study. For example, the gene ADRB2, is associated to several multimorbidities: COPD with cardiovascular diseases (atrial fibrillation, heart failure, ischemic heart disease and stroke), and with diabetes, lung cancer and obesity. Interestingly, the protein encoded by this gene has been shown to participate in smooth muscle relaxation in bronchi resulting in a facilitated respiration [26]; in blood vessels, it dilates coronary and skeletal muscle arteries [27] and it also contributes to insulin secretion from pancreas [28]. ADRB2 has been the focus of pharmacogenomics studies in COPD: several polymorphisms of this gene have been shown to influence patient response to bronchodilators treatment [29]. A recent study showed a significant correlation between the R16G polymorphism with COPD severity in term of FEV1 (forced expiratory volume in 1 second) in a Greek population [30]. Interestingly, the R16G–Q27E haplotype has been shown to be associated with glucose tolerance and insulin sensitivity in obese postmenopausal women [27]. Thus, this polymorphism might influence both bronchodilator response and glucose tolerance in specific patient populations, constituting a potential link to explain the association between COPD and diabetes.

Another study in a Danish population showed that the ADRB2 T164I polymorphism is associated with a reduced lung function and an increased risk of COPD in the general population [31]. This polymorphism is also associated with increased blood pressure, higher frequency of hypertension and increased risk of ischemic heart disease amongst women in the general population [26]. Thus, the ADRB2 T164I polymorphism could explain the association between COPD and cardiovascular diseases.

Further analysis using haplotype frequencies indicated that haplotypes G16-Q27-I164 and (non-G16-Q27)-T164, as compared to the reference haplotype G16-Q27-T164, were significantly associated with a decreased risk of myocardial infarction [32]. In summary, the ADRB2 gene represents an interesting candidate to explain the multimorbidity of COPD with diabetes, obesity and ischemic heart disease.

To identify the most relevant biological functions of the genes and proteins shared by the COPD diseasome, we performed a functional enrichment analysis with biological pathways from Reactome on the set of common proteins, as detailed in the Methods section. Significant annotations were those with q-values ≤ 0.05, and results are visualized here as heat-maps. We observed that “Innate immune response” pathways were enriched in almost all of the multimorbidity pairs studied, particularly in those of COPD with diabetes, osteoporosis, lung cancer, heart failure or ischemic heart disease (Additional file 3: Figure S2 in the Supporting information). “Adaptive immune response” pathways showed a similar pattern: they were also enriched in all the pairs with the exception of those of COPD with metabolic syndrome, anemia, sleep apnea and cachexia (Additional file 4: Figure S3 in the online data supplement). Likewise, with the exception of the pair pulmonary hypertension-COPD, “Apoptosis” pathways and related signaling events, such as “TNF signaling” or “Death receptor signaling” were also involved in most multimorbidity pairs (Additional file 5: Figure S4 in the online data supplement). Interestingly, proteins involved in the “Nitric oxide metabolism” pathway were associated with COPD multimorbidities involving the cardiovascular system (atrial fibrillation, heart failure, ischemic heart disease and stroke), metabolic diseases (diabetes, metabolic syndrome), osteoporosis, pulmonary hypertension, lung cancer and depression (Additional file 6: Figure S5 in the online data supplement). “Hemostasis” pathways, including “Platelet activation, signaling and aggregation” were also enriched in all the pairs studied (Figure 5). Finally, “Cell cycle” pathways were enriched in lung cancer-COPD, diabetes-COPD and, surprisingly, in depression-COPD (Additional file 7: Figure S6 in the online data supplement).

To explore the degree of similarity between two COPD multimorbidities (for instance lung cancer-COPD and diabetes-COPD) at the level of biological pathways, we used the Jaccard coefficient (see Methods). We found the highest Jaccard coefficient values among pairs of COPD and several CVDs, including heart failure, ischemic heart disease and stroke, which share approximately 50% of the involved pathways (Figure 6). Likewise, the level of pathway coincidence was also considerable between pairs of CVD-COPD and pairs of COPD with nutritional and metabolic diseases. More surprisingly was the observation of a relatively high Jaccard coefficient between CVD-COPD pairs and osteoporosis-COPD, as well as that lung cancer-COPD and diabetes-COPD that showed a pathway similarity of 48% (lung cancer and diabetes without taking into account COPD showed only 35% of pathway similarity). Albeit to a lesser degree, nutritional and metabolic diseases-COPD and musculoskeletal diseases-COPD groups present pathway similarities as well. Other interesting associations that emerged from this analysis included the overlap between depression-COPD and diabetes-COPD, and between depression-COPD and heart failure-COPD. Finally, it was also of interest to note that the biological pathways involved in some comorbid disease-COPD pairs were almost unrelated to those of others (for instance, anemia-COPD, cachexia-COPD, sleep-apnea-COPD or polycythemia-COPD) (Figure 6). All in all, these observations highlight interesting associations between COPD multimorbidities at the biological pathway level.

According to the CTD [33], the 106 tobacco smoke chemicals present in the tobacco smoke [22],[23] target an average of 69% of the proteins shared by COPD and the multimorbidities shown in Figure 4 (see complete data set on Additional file 8: Table S2 in the online data supplement). For instance, 468 out of the 759 (61.7%) proteins shared by COPD and lung cancer are known targets of at least one compound present in tobacco smoke. This proportion was even higher (72%) in the case of stroke with COPD. This is not surprisingly since tobacco smoke is a main risk factor of COPD, CVD and lung cancer. Yet, even in multimorbidities that, at first glance, might seem unrelated to smoking, such as metabolic syndrome, diabetes or anemia, we also found a high percentage of shared proteins that are established targets of tobacco smoke compounds in the CTD, which provides potential candidates to investigate the role of tobacco smoke compounds in COPD multimorbidities.

The novelty, as well as the unbiased and integrative nature of our scientific approach, is a clear strength of our analysis. However, some limitations are worth noting. The most relevant one is that our analysis depends on the scientific information that is available and recorded in public databases, such as Medline, Reactome and other repositories of biological data. Further, due to limitations in database development and curation, many of such databases are likely to be incomplete [39]. Hence, the results of our analysis should be mostly regarded as hypothesis generators requiring experimental validation.