Protein interaction networks provide insight into fetal origins of chronic obstructive pulmonary disease

We used published results from a fetal lung DNA methylation data set and COPD DNA methylation and expression data sets [36‐38].

The fetal lung DNA samples included 78 samples that passed the quality control measures [36]. Methylation in smoke-exposed was compared to unexposed fetal lung samples and were considered nominally significant at a p-value cut off of 0.05. The fetal lung DNA samples were isolated from discarded tissue from 8–18 weeks of gestation. The samples were anonymized at study entry at the Laboratory of Developmental Biology, University of Washington, Seattle, WA, USA.

Genome-wide methylation assay was performed using 750 ng of bisulfite-treated DNA per sample using the Infinium HumanMethylation450 BeadChip array (Illumina, San Diego, CA, USA), according to manufacturer’s recommended protocol. Data were available for gestational age, fetal sex, and cotinine levels. Sex was verified using X chromosome methylation. IUS exposure was inferred by measuring placental cotinine concentrations. Exposure was treated as a continuous and dichotomous variable, with levels of cotinine ≤ 7.5 ng/g considered as unexposed (control group) and levels of cotinine > 7.5 ng/g as exposed. Published results were used from site based differential methylation analysis from limma (version 3.37.7) [39] adjusting for age, sex, sample plate, and sentrix position. DM CpG sites were nominally significant at a p-value cut off of 0.05 and mapped to genes using Human Genome build: GRCh37/hg19 annotation.

Genome-wide methylation assay was performed using 750 ng of bisulfite-treated DNA per sample using the Infinium HumanMethylation450 BeadChip array (Illumina, San Diego, CA, USA) and gene expression was assayed using the Illumina HumanHT-12 Bead Chips [37, 38]. CpG sites were mapped to genes using Human Genome build: GRCh37/hg19 annotation.

The study included lung tissue samples from 114 COPD cases (avg. age 63.4, 60% males, all former smokers, quit smoking 84.7 months before on avg., FEV 1% predicted 26.3 avg.) and 46 control smokers with normal lung function (avg. age 65.3, 29% males, all former smokers, quit smoking 181 months before on avg., FEV 1% predicted 98.1 avg.).

Published results were used fromsite based differential methylation and gene-based expression analyses performed using limma (version 3.37.7) [39]. Previously published results [37, 38] were included at a p-value cut off of 0.05. CpG sites were mapped to genes using Human Genome build: GRCh37/hg19 annotation.

In order to find meaningful connected components, a PPI of decent size and non-sparsity is required. The predictive power of the connectivity significance increases as the PPI becomes more complete [41]. We used the HumanNet-FN [35] PPI (downloaded April 2019 https://​www.​inetbio.​org/​humannet) which includes co-functional links (given by co-essentiality, co-expression, pathway database, protein domain profile associations, gene neighborhood, and phylogenetic profile association) and protein–protein interactions (given by high-throughput assays and literature curated interactions). The network consists of 17,247 genes, which are connected by 371,502 undirected edges (where 118,012 are physical, 213,003 functional, and 39,587 are physical and functional interactions). The largest connected component (LCC) of the PPI consists of 17,191 genes which are linked to each other by 371,464 edges.

An overview of the data sets and their LCCs in the PPI can be found in Table 1.

The method used here is an extension of the work of Wang et al [42] which selects all nominally-significant genes (p-value < 0.05) and then uses fold change values for ranking genes. The framework identifies exposure or disease modules by agglomerating genes based on their statistical significance within their respective study.

Our approach here is similar, except that it considers all genes of the data set (not only nominally-significant genes), ranking them according to their p-value (rather than fold change), from the most significant to the least significant. The remaining steps are the same as in [42]. First, different thresholds for the p-values are given. Next, for each threshold the LCC is identified which is given by all genes which have a p-value lower than the threshold. With increasing p-value thresholds the sizes of the LCCs increase. The sizes of the LCCs are then compared against random expectation and a z-score is computed to indicate their significance. Thus, we obtain a p-value threshold vs. z-score plot which is used to determine the module. The module is the LCC with a z-score above 1.6 and of a size which is in general considered to be a reasonable size for a module (30–100) containing genes which have relatively small p-values. If several LCCs match these criteria we choose the one with the highest z-score. Thus, the method ensures that the genes which can be most strongly associated with a phenotype of interest are preferentially added to the module while maintaining significant module connectivity. We provide a detailed method description in Additional file 1 (section “Computation of the modules”).

We identified one module for each methylation set (fetal lung and adult COPD) and one for the COPD gene expression set. Additionally, we computed two modules for the 502 genes found in the fetal lung and COPD sets. Here, a module was computed using the p-values given by the fetal lung methylation data set and another one was computed using the p-values given by the COPD methylation data set (Additional file 1 section “Computation of the modules using genes which are significantly enriched in both methylation data sets” and “Modules computed using genes which are significantly enriched in both methylation data sets”).

To study the topological robustness of the modules, we evaluated whether highlighted module genes form significantly connected components in five different PPIs (BioGRID [43], STRING [44], Hint [45], PPI2016 [46], and BioPlex [47]). To do so, we first identify the LCC given by the modules’ genes in the other PPIs and next compared this size against random expectation. All modules form significantly connected component in all five PPIs except for the COPD methylation module in the STRING PPI. These results show that the modules (and the method) are robust irrespective of the choice of PPI (Additional file 1 section “Robustness” and Additional file 2: Table S1).

In order to identify genes previously associated with COPD we used the database DisGeNet [48]. We entered each gene individually and filtered the “Summary of Gene Disease Association’s” results for “Disease Classes” containing “Respiratory Tract Diseases”.

We performed enrichment analyses on different sets of genes given by the computed modules and their connections to the other modules. For all analyses we used g:Profiler [49] (accessed May 2020) using the 17,190 genes in the LCC in the HumanNet-FN (Additional file 3: Table S2) as background and the default parameters otherwise. We considered a pathway as significantly enriched with a p-value < 0.05. We performed an enrichment analysis for each set of genes in each module and for each set of interactors.

The set of 5175 genes in the fetal lung methylation data set produced an IUS-exposure module of 50 genes (Table 2). We found that 7 of the 50 genes (14%) (hypergeometric p-value = 0.04) have been related to COPD (Fig. 1b, Additional file 1 section “Fetal lung methylation module” and Table 3).

All results, including the Gene Disease Association score can be found in the Table 3 and Additional file 4: Table S3. Additionally, we looked for associations of genes to COPD according to GWAS study using the study of Sakornsakolpat et al. [52].

The COPD disease module given by the 1217 genes in the COPD methylation data set (adj. P-value < 0.05) [37] consists of 37 genes (Table 2), and 4 (11%) have prior associations to COPD (hypergeometric p-value = 0.15) (Fig. 1c, Additional file 1 section “COPD methylation disease module” and Table 3).

There are 204 genes significantly differentially expressed in the adult COPD gene expression data set (adj. p-value < 0.05) [37] and the resulting disease module consists of 64 genes (Fig. 1d, Table 2). Twelve genes of the module (19%) have prior associations with COPD (hypergeometric p-value = 0.001) (Additional file 1 section “COPD expression module” and Table 3).

The three modules support genomic links between IUS-exposure and COPD in adults. The methylation modules for fetal and adult lung do not overlap and the fetal lung methylation module and the COPD expression module have only one gene in common (BCL11A). Therefore, we focused using our method to explore genes connecting the fetal lung IUS exposure and adult COPD PPI modules. Both COPD disease modules contain genes which are directly connected to genes of the fetal lung methylation module in the HumanNet-FN (Fig. 1e). The number of edges connecting these modules is higher than expected by chance (p-value < 1e−05) (Additional file 1 section “Connectivity between the modules”); most edges (196 out of 286, 69%) connecting the modules with each other are functional. In total there are 66 genes which connect one module with another and we call these genes interactors. Twenty-seven interactors are members of the fetal lung methylation module, of which 13 connect to the COPD methylation disease module and 23 to the COPD expression disease module (9 genes are connected to both modules) (Table 2). Fifteen genes of the COPD methylation disease module and 24 genes of the COPD expression disease module connect to the fetal lung methylation module (Figs. 1e and 3a, Tables 3 and Additional file 3: S2). Genes with prior known associations to COPD in the literature are well connected (z-score = 8.1, p-value = 1.4e−5) (Additional file 1 section “Connectivity of the genes which can be associated to asthma and/or COPD”), especially between the three modules, with predominant functional edges (hypergeometric p-value = 1.4e−05) (Fig. 2). There are in total 21 genes in the modules which can be associated with COPD. Not all of them are connected to each other, but the largest connected component contains 13 genes (Table 3). Half of the 24 interactors of the COPD expression module which are connected to the fetal lung methylation module are up-regulated while the other half is down-regulated. Sixteen out of the 23 interactors in the COPD expression module connected to the COPD methylation module are down-regulated (Additional file 5: Table S4).

The interactors, as linking genes, are of potential interest since we hypothesize that these may capture genomic trajectories between perturbations in lung tissue during fetal development and COPD in adulthood. Therefore, the 66 interactor genes were subjected to pathway enrichment analysis to identify perturbed pathways that may mark susceptibility to COPD.

We performed enrichment analyses on seven gene sets given by the modules and their connections (Figs. 1e, 3a), using KEGG [53], and the LCC of the HumanNet-FN as background.The results of the enrichment analyses can be found in Fig. 3 and 4, as well as in Additional file 6: Table S5. First, we performed three enrichment analyses using the whole set of genes of each module, including the fetal lung methylation module (50 genes), the COPD methylation module (37 genes), and the COPD expression module (64 genes) (Table 2).

Next, we performed an enrichment analysis for each set of interactors: the set of genes from the fetal lung methylation module which are connected to the COPD methylation module (14 genes) and the set of genes from the fetal lung methylation module which are connected to the COPD expression module (23 genes), the set of genes from the COPD methylation module which are connected to the fetal lung methylation module (15 genes), and the genes from the COPD expression module which are connected to the fetal lung methylation module (24 genes).

All significantly enriched pathways (adj. p-value < 0.05) for at least three sets of the genes defined above are listed in the table in Fig. 3b (see Additional file 6: Table S5 for more details). The pathway which was significantly enriched for most gene sets (four out of seven gene sets) was the AGE-RAGE pathway, followed by the Focal-Adhesion pathway.