A Multiomics Approach to Identify Genes Associated with Childhood Asthma Risk and Morbidity

Abstract

Childhood asthma is a complex disease. In this study, we aim to identify genes associated with childhood asthma through a multiomics “vertical” approach that integrates multiple analytical steps using linear and logistic regression models. In a case–control study of childhood asthma in Puerto Ricans (n = 1,127), we used adjusted linear or logistic regression models to evaluate associations between several analytical steps of omics data, including genome-wide (GW) genotype data, GW methylation, GW expression profiling, cytokine levels, asthma-intermediate phenotypes, and asthma status. At each point, only the top genes/single-nucleotide polymorphisms/probes/cytokines were carried forward for subsequent analysis. In step 1, asthma modified the gene expression–protein level association for 1,645 genes; pathway analysis showed an enrichment of these genes in the cytokine signaling system (n = 269 genes). In steps 2–3, expression levels of 40 genes were associated with intermediate phenotypes (asthma onset age, forced expiratory volume in 1 second, exacerbations, eosinophil counts, and skin test reactivity); of those, methylation of seven genes was also associated with asthma. Of these seven candidate genes, IL5RA was also significant in analytical steps 4–8. We then measured plasma IL-5 receptor α levels, which were associated with asthma age of onset and moderate–severe exacerbations. In addition, in silico database analysis showed that several of our identified IL5RA single-nucleotide polymorphisms are associated with transcription factors related to asthma and atopy. This approach integrates several analytical steps and is able to identify biologically relevant asthma-related genes, such as IL5RA. It differs from other methods that rely on complex statistical models with various assumptions.

Clinical Relevance

This study integrates several pathogenic steps and “omics” data to identify biologically relevant asthma-related genes, such as IL-5 receptor α. Unlike other methods that rely on complex statistical models with various assumptions, each step in our approach uses linear or logistic regression and is simple to interpret.

Asthma is the most common noncommunicable chronic disease of childhood (1) and affects over 6 million children in the United States, leading to over 3.5 million exacerbations, 135,000 hospitalizations, and approximately 14 million missed school days each year (2). It is a multifactorial disease with heritability as high as 80–90% (3, 4). To date, several dozen asthma-related genes have been identified (5), yet single-nucleotide polymorphisms (SNPs) in such genes account for less than half of the disease’s heritability.

Our ability to identify novel genetic polymorphisms associated with disease has increased exponentially with the use of microarrays and now high-throughput technologies, such as sequencing. However, this steep increase in the number of markers analyzed has also introduced new problems in terms of statistical analysis and biological interpretation of novel results (6). Strategies used to deal with these issues suffer from their own problems: mathematical corrections for multiple comparisons can be excessively conservative and may discard associations that could be biologically important; different methods rely on various assumptions that are not necessarily generalizable to other circumstances, et cetera (7). This is particularly important for analysis of multiomics data (7). Conversely, attempts at replicating findings in independent cohorts presume that the populations are similar enough or that the specific polymorphisms are universally relevant.

The pathogenesis of complex diseases, such as asthma, involves several steps: genetic variants are subject to epigenetic regulation, which helps determine the level of gene expression; expressed genes are, in turn, translated into proteins, which may have direct effects on disease “intermediate phenotypes”; and these intermediate processes ultimately alter an individual’s risk of disease or disease severity. In this study, we propose a novel approach that aims to identify genes and pathways that are relevant to asthma by integrating several analytical steps that are based on the conceptual framework of disease pathogenesis. Furthermore, we aimed to do so through a “vertical” approach that uses existing data in the same population to sequentially narrow our analytical focus, thereby maximizing efficiency, minimizing multiple comparisons, and avoiding complex mathematical models built on numerous assumptions.

Materials and Methods

Study Population

As part of a case–control study of childhood asthma in Puerto Ricans (Hartford and Puerto Rico; HPR), we recruited 1,127 children (618 with asthma [cases] and 509 without asthma [control subjects]). The details of subject recruitment have been published previously (8–10). In brief, from September 2003 to July 2008, 449 Puerto Rican children were recruited from elementary/middle schools in Hartford, Connecticut, and from March 2009 to June 2010, 678 Puerto Rican children living in San Juan, Puerto Rico, were recruited using a multistage probability sample design. A household was eligible if one or more resident was a child 6–14 years old. Participants had to have four Puerto Rican grandparents and had to be living in the same household for 1 year or longer. In households with more than one eligible child, one child was randomly selected for screening. There were no significant differences in age, sex, or area of residence between eligible children who did and did not participate at either study site. We selected, as cases, children with physician-diagnosed asthma and at least one episode of wheeze in the prior year, and, as control subjects, children with no physician-diagnosed asthma and no wheeze in the prior year. Written parental consent was obtained for participating children, from whom written assent was also obtained. The study was approved by the Institutional Review Boards of Connecticut Children’s Medical Center (Hartford, CT), the University of Puerto Rico (San Juan, PR), Brigham and Women’s Hospital (Boston, MA), and the University of Pittsburgh (Pittsburgh, PA).

Study Procedures

Study participants completed a protocol that included questionnaires, spirometry, blood sample collection, and skin test reactivity (STR), as previously described (11). In brief, spirometry was conducted according to American Thoracic Society criteria modified for children (12); spirometry was repeated 15 minutes after administration of 200 μg of albuterol metered-dose inhaler. STR included histamine (positive control), saline solution (negative control), and allergen extracts from house dust mite mix (Dermatophagoides pteronyssinus and Dermatophagoides farinae), German cockroach (Blattella germanica), cat dander, dog dander, mixed grass pollen, mugwort sage, ragweed, mixed tree pollen, mold mix, Alternaria tenuis, and mouse extracts. Peripheral blood eosinophil count was measured by Coulter counter techniques and then log₁₀ transformed for analysis. In addition to asthma status, five intermediate phenotypes were selected a priori for analysis: age of asthma onset; moderate–severe asthma exacerbations (defined by EPR-3 criteria [13]); post-bronchodilator forced expiratory volume in 1 second (FEV₁); eosinophil counts; and the total number of STR⁺.

Cytokine Plasma Levels

For a previously published analysis (10), 14 Th17 pathway-related cytokines (IL-1b, IL-4, IL-6, IL-10, IL-17A, IL-17F, IL-21, IL-22, IL-23, IL-25, IL-31, IL-33, IFN-γ, and TNF-α) were measured in plasma using the Bio-Plex Pro Human Th17 cytokine panel (Bio-Rad Laboratories, Hercules, CA). Assays were designed on magnetic beads in a capture-sandwich immunoassay format. Undetectable cytokine levels were assigned a constant (half the lowest limit of detection). Cytokine data were available for 578 participants. All cytokine levels were log₁₀ transformed for data analysis.

Genome-Wide Genotyping

Genome-wide (GW) genotyping (GWG) was performed using the Illumina HumanOmni2.5 BeadChip platform (Illumina, San Diego, CA), which measures approximately 2.5 million SNPs, as previously described (14, 15). Subjects with a call rate less than 95% were removed from the analysis. SNPs were removed if they were not in Hardy–Weinberg equilibrium (P < 10⁻⁶) in control subjects, had minor allele frequency lower than 1%, or had a failure rate greater than 5%. After all subject and marker quality control steps, approximately 1.83 million SNPs for 951 HPR participants were included in the analysis.

GW White Blood Cell DNA Methylation

High-quality white blood cell (WBC) DNA (750 ng) was available from 340 HPR participants. After bisulfite conversion using EZ-DNA methylation kits (Zymo Research, Irvine, CA), DNA (200 ng) was used for GW WBC DNA methylation (GWM) analysis using the HumanMethylation 450K BeadChip (Illumina), as previously described (16). Methylation data were read using the methylumi R package (version 3.0.1; https://www.r-project.org/about.html), and β values for each CpG were calculated as: β = M / (M + U + α), where M and U represent methylated and unmethylated signal intensities at the specific site, and α is an arbitrary offset (usually 100) intended to stabilize β values where fluorescent intensities are low (16). Raw methylation data were normalized with the SWAN (subset-quantile within array normalization) method using the minfi R package (16, 17).

GW Whole-Blood Gene Expression Profiling

RNA was available from 141 HPR participants. After extraction from whole blood using PAXgene blood miRNA kits (Qiagen, Valencia, CA), globin transcripts were then depleted using GLOBINclear kit (Life Technologies, Carlsbard, CA). RNA quality and concentration were determined using Agilent RNA 6,000 Nano kit (Agilent Technologies, Santa Clara, CA). GW whole-blood gene expression (GWE) profiling was measured in 141 whole-blood globin-cleaned RNA samples using the HumanHT-12 v4 Expression BeadChip (Illumina). Background subtraction and quantile normalization were performed using the lumi R package, as previously described (15). Probes with greater than 70% absent points among all 141 samples were excluded from the downstream analysis. A total of approximately 15,000 probes was included in the final analysis. Please refer to Figure E1 in the online supplement for overlap between GWG, GWM, and GWE availability for the cohort.

Principal Component Analysis

Principal component (PC) analysis was performed using genotyping data, methylation data, and expression data separately. PCs can capture different sources of variation due to global ancestry, population stratification, unmeasured environmental factors, or technical factors. The first two PCs derived from GWG data, the first 10 PCs from GWE data, and the first two PCs from GWM data were included as covariates in the respective analytic steps.

Analytical Approach

Please see the online supplement for further details. We performed our analysis in several steps, working under a “vertical” approach that follows the conceptual framework of pathogenic pathway from genotype to disease (Figure 1)—including genetics (GWG), epigenetics (GWM), genomics (GWE), proteomics (cytokine levels), phenomics (asthma-related intermediate phenotypes), and disease status (asthma versus control):

• 1.Modification of gene expression versus protein level by asthma status: our initial hypothesis was that gene expression and protein level would be most closely associated; furthermore, we hypothesized that genes and pathways implicated in asthma risk or morbidity would be differentially activated between cases and control subjects. Therefore, in this first step, we evaluated the association between GWE and cytokine levels by asthma status using the general model: P_i ∼ A + E_j + A⋅E_j + covariates, where P_i is the plasma level of cytokine, i, A is the participant’s asthma status, E_j is the expression level of gene, j, and A⋅E_j is the interaction term. Next, we evaluated genes the association of which with protein level differed by asthma status (interaction P < 0.01) in a gene-set enrichment analysis using Fisher’s exact test (Reactome and KEGG databases), and selected the top enriched pathway(s) as candidate(s) for subsequent steps.
• 2.Gene expression versus intermediate phenotypes: we assessed the expression levels of the genes carried forward from step 1 versus five asthma-related intermediate phenotypes (age of asthma onset, asthma moderate/severe exacerbations, baseline FEV₁, peripheral blood eosinophil counts, and positive allergy skin testing) using linear or logistic regression as appropriate. From this analysis, we selected the genes with the 20 lowest P values plus those with P less than 0.01 in at least three of five intermediate phenotypes. In addition, we included one gene (IL18) that was nominally significant in step 1, but nonsignificant in step 2 as a “negative control,” hypothesizing that it would not be consistently significant in subsequent analytical steps.
• 3.Methylation versus asthma status: next, we evaluated the association between DNA methylation levels at the CpG sites of the genes carried forward from step 2 and asthma status using regression modeling. We carried forward genes with false discovery rate–adjusted P less than 0.05.
• 4.Genotype versus asthma status: we then evaluated the association between all SNPs (from GWG) in the top genes from step 3 and asthma status using regression modeling.
• 5.Expression quantitative trait loci (eQTL) analysis: simultaneously, we evaluated SNPs that were associated with the expression levels of the top genes from step 3, based on result from our recent eQTL analysis (15).
• 6.Genotype versus protein level: we included all SNPs that were significant in steps 4 or 5 and evaluated their association with the plasma levels of the corresponding cytokines from step 1.
• 7.Genotype versus intermediate phenotypes: we included all SNPs that were nominally significant in step 6 and evaluated the association between the SNPs and the corresponding intermediate phenotypes from step 2.
• 8.Protein level versus intermediate phenotypes: in the final step, we analyzed the association between the protein levels that showed nominally significant associations in step 1 and the intermediate phenotypes from steps 2 and 7.

Figure 1.

Pathogenic and analytical steps. Pathogenic steps from genotype to disease (left) and analytical steps (right) performed in the current study. Shown are the numbers of genes, probes, or single-nucleotide polymorphisms (SNPs) included in each sequential step, as well as the generic analytical model used. All models were adjusted for pertinent covariates (not shown in the figure for simplicity).

IL-5 receptor α and in silico transcription factor analysis

We measured IL-5 receptor α (IL-5Rα) plasma levels (the protein for which gene IL5RA codes) in a subset of participants (n = 130) and evaluated the association between IL-5Rα levels and asthma phenotypes. In addition, we assessed the effects of IL5RA SNPs on IL-5Rα levels and their modification by asthma status. All models were adjusted for age, sex, parental asthma, and household income. We also evaluated the regulatory potential of IL5RA SNPs using is-rSNP (18), software designed to predict whether an SNP has a regulatory effect on transcription factor binding sites in silico.

Results

Modification of Gene Expression versus Protein Level by Asthma Status

Characteristics of study participants are shown in Table 1. Asthma modified the gene–protein association in a total of 1,645 genes (interaction P < 0.01); of the 14 cytokines tested, the main associations were with IL-10, IL-17A, IL-17F, IL-22, and IL-23. Results from gene-set enrichment analysis for those 1,645 genes showed that the Reactome cytokine signaling system (n = 269 genes; see Figure 2) was significantly enriched for those genes and cytokines (see Table E1). Thus, these 269 genes were carried forward.

Table 1.

Characteristics of Study Participants

	Children with GWG	Children with GWM	Children with GWE
No. of children	948	309	138
Age, yr	10.1 (2.7)	8.6 (1.9)*	10.4 (2.7)
Male sex, %	51.6	57.3	56.5
FEV1, L	1.95 (0.71)	1.60 (0.51)	1.98 (0.71)
Eosinophils, cells/mm3	239 (133–426)	389 (202–599)*	340 (181–554)*
STR+, %	60.9	74.1	76.9
Asthma, %	54.9	56.6	50.7
Age of onset, yr†	2.4 (2.5)	2.2 (2.0)	2.6 (2.4)
Exacerbation, %†	42.2	58.9	55.7

Definition of abbreviations: GWE, genome-wide expression; GWG, genome-wide genotypic; GWM, genome-wide methylation; STR⁺, positive skin test reactivity.

Shown are mean (SD) or median (interquartile range) for continuous variables, and percent for dichotomous variables.

^*P < 0.05 compared to data from children with GWG, but without GWM or GWE.

^†Among children with asthma only.

Figure 2.

Venn diagram showing number of genes in top pathways from gene-set enrichment analysis. Reactome pathway “cytokine signaling in the immune system” (n = 269 genes) and the contained pathways were highly enriched for genes significant in analytical step 1 (modification of gene expression versus cytokine level by asthma status). The next encompassing pathway (“immune system”) was deemed too broad for further analysis.

Gene Expression versus Intermediate Phenotypes

Age of onset showed nominally significant associations with n = 60 of 269 genes (22%); exacerbations with n = 42 genes (16%); FEV₁ with n = 40 (15%), eosinophils with n = 32 (12%); and SPT⁺ with n = 66 (25%)—all significantly more than expected by chance. From this step, we selected 40 genes that were among the top P values and those consistently associated with more than two intermediate phenotypes. In addition, IL18 was included as a negative control. Table E2 shows the detailed results from this step.

Methylation versus Asthma Status

A total of 7 of 41 genes (17%) had DNA methylation levels that were associated with asthma status: Conserved helix-loop-helix ubiquitous kinase (CHUK), IL-1 receptor-2 (IL1R2), IL-1 receptor antagonist (IL1RN), IL-5Rα (IL5RA), IFN regulatory factor 7 (IRF7), tyrosine-protein kinase Lyn (LYN), and suppressor of cytokine signaling 2 (SOCS2). IL5RA had the most significant P value and the most differentially methylated CpG sites. Table 2 shows the summary findings for these seven genes (for all analytical steps).

Table 2.

Stepwise Analytical Approach Summary for Seven Final Genes

Step	CHUK	IL1R2	IL1RN	IL5RA	IRF7	LYN	SOCS2	IL18	Comments
1	0.03	0.007	0.01	0.007	0.038	0.093	0.10	0.04	Top interaction P value
	IL-23	IL-17A	IL-17A	IL-17F	IL-17A	IL-17A	IL-23	IL-23	Associated cytokine(s)
2	n/s	0.002	0.007	n/s	0.009	0.027	0.022	n/s	P value for exacerbations
	0.041	0.048	0.006	n/s	n/s	0.009	0.005	n/s	Age of onset of asthma
	0.042	n/s	n/s	n/s	n/s	n/s	0.023	n/s	FEV1
	0.031	0.03	n/s	0.046	0.042	0.015	0.011	n/s	STR+
	n/s	n/s	0.014	6.3 × 10−12	n/s	n/s	0.042	n/s	Eosinophils
3	0.042	0.039	0.031	0.012	0.049	0.042	0.042	0.02	Top FDR P value
	1	1	1	4	1	1	1	2	Number of FDR P < 0.05 sites
4	n/s	n/s	n/s	0.019	n/s	n/s	n/s	n/s	Top FDR P value
	—	—	—	rs17878498	—	—	—	—	Associated SNP(s)
5	0.024	0.052	0.026	0.0055	0.015	0.014	0.0016	0.009	Top FDR P value
	4	0	5	17	4	25	56	10	Number of FDR P < 0.05 SNPs
6	n/s	n/s	n/s	0.024	0.022	0.025	n/s	n/s	Top P value
	—	—	—			6	—	—	Number of P < 0.05 SNPs
7	n/s	n/s	n/s	0.0507	0.01	0.049	0.005	n/s	Top P value
	—	—	—	Eosinophils	Exacerbations	Total STR+	Onset age	—	Associated phenotype(s)
8	n/s	n/s	n/s	0.013	n/s	n/s	n/s	n/s	Top P value
	—	—	—	Eosinophils	—	—	—	—	Associated phenotype(s)

Definition of abbreviations: FDR, false discovery rate; n/s, not statistically significant; SNP, single-nucleotide polymorphism; STR⁺, positive skin test reactivity.

Final candidate genes: conserved helix-loop-helix ubiquitous kinase (CHUK); IFN regulatory factor 7 (IRF7); IL-1 receptor-2 (IL1R2); IL-1 receptor antagonist (IL1RN); IL-5 receptor α (IL5RA); tyrosine-protein kinase Lyn (LYN); and suppressor of cytokine signaling 2 (SOCS2).

Bold typeface denotes P < 0.05.

SNP versus eQTL

We evaluated all SNPs associated with expression of the seven selected genes. Table 2 shows the false discovery rate–adjusted P values as well as the corresponding SNPs. SNP rs17878498 was associated with expression of IL5RA; it is located approximately 870 bases upstream from IL5RA in chromosome 3p26.6, and is part of the binding site motif for transcription factor SPI1, which is also involved in IL-4 signaling and glucocorticoid receptor regulation.

SNP versus Asthma Status

Table 2 shows the top P value associated with asthma status for the seven selected genes, as well as the total number of SNPs with P less than 0.05 for each gene. Three genes (IL5RA, LYN, and SOCS2) had the top P values and/or the most SNP associations (17, 25, and 56, respectively).

SNP versus Cytokine Level

We included 134 SNPs in the 7 selected genes. Three genes (IL5RA, IRF7, and LYN) had SNPs that were significantly associated with the cytokine with which there was differential expression–protein association by asthma status (step 1). IL5RA SNP rs6808454 was associated with IL-17F plasma levels (P = 0.024), IRF7 SNP rs143197172 with IL-17A (P = 0.022), and five SNPs in high linkage desequilibrium in LYN SNP with IL-17A (P = 0.025). In addition, each gene had other SNPs with P less than 0.05 (five total SNPs in IL5RA, two in IRF7, and five in LYN).

SNP versus Intermediate Phenotype

Among SNPs that were significant in step 6, IL5RA SNP rs6773701 was associated with eosinophil counts (P = 0.05), IRF7 SNP rs142972371 with exacerbations (P = 0.016), and LYN SNP rs112735328 was associated with the total number of STR⁺ (P = 0.009).

Cytokine Level versus Intermediate Phenotype

Finally, in step 8, IL-17F (associated with IL5RA) was significantly associated with higher eosinophil counts (β = 0.30; 95% confidence interval [CI] = 0.06–0.54; P = 0.013), and also showed borderline association with lower FEV₁ (β = −248 ml; 95% CI = −511 ml to +16 ml; P = 0.067).

Of the seven selected genes, IL5RA was significant in all eight analytical steps. IRF7 was significant in six steps, and the other five genes were significant in four to five steps. Contrarily, IL18, which was selected as a “negative control” because it did not pass step 2, was only significant in three out of eight steps.

IL-5Rα plasma levels

We measured IL-5Rα plasma levels in a subset of participants: median IL-5Rα level was 0.08 pg/ml (minimum, 0.058 pg/ml; maximum, 3.61 pg/ml). In our adjusted analysis, children who had moderate–severe asthma exacerbations had lower IL-5Rα levels (−1.98 pg/ml [95% CI = −3.81 to −1.03]; P = 0.044) and children with earlier age of onset had higher IL-5Rα (+1.17 pg/ml [95% CI = 1.01–1.36]; P = 0.033). In addition, the effects of five IL5RA SNPs on IL-5Rα levels were modified by asthma status (Figure E1). Table 3 shows the summary of all SNPs in IL5RA that showed nominally significant associations in our stepwise analysis.

Table 3.

IL5RA Single-Nucleotide Polymorphisms with Significant Associations in the Stepwise Analysis

Step 4 Gene Expression ∼ SNP		Step 5 Asthma Status ∼ SNP		Step 6 Cytokine (IL17F) ∼ SNP		Step 7 Phenotypes ∼ SNP		IL5-Rα Level ∼ SNP⋅Asthma
SNP	P Value	SNP	P Value	SNP	P Value	SNP	P Value	SNP	P Value
rs17878498	0.018	rs3804790	0.005	rs6808454	0.024	rs6773701	0.0507	rs340832	0.043
									0.034†
		rs11294168	0.007	rs6773701	0.034			rs340833	0.10
									0.009†
		rs1153461	0.016	rs71058675	0.045			rs340807	0.18
									0.059†
		rs35439082	0.017	rs2600027	0.046			rs340809	0.051
									0.010†
		rs3856847	0.025	rs3804790	0.049			rs3792424	0.073
									0.029†
		rs1153462	0.027
		rs7644398	0.029
		rs71058675	0.029
		rs35418165	0.031
		rs62230291	0.035
		rs17887062	0.036
		rs17879465	0.041
		rs6808454	0.042
		rs142128948	0.043
		rs150254181	0.045
		rs2600027	0.049
		rs6773701	0.049

Definition of abbreviation: SNP, single-nucleotide polymorphism.

All SNPs sorted by P value. SNPs significant in more than one step are shown bold typeface.

^†Interaction term (IL-5 receptor α ∼ SNP + SNP⋅asthma + covariates).

In silico transcription factor analysis

Finally, we evaluated the regulatory potential of IL5RA SNPs using is-rSNP (18). Table E3 shows all results with adjusted P less than 0.05. Among the top transcription factors associated with these IL5RA SNPs, several have been associated with asthma or atopy, or other lung processes and diseases, including hypoxia-inducible factor 1α and 1β (19–21), serum response factor (22, 23), upstream transcription factor 1 (24), CCAAT-enhancer binding protein β (25), peroxisome proliferator–activated receptor α (26), signal transducer and activator of transcription 3 (27), Runt-related transcription factor 1 (28), et cetera.

Alternative/confirmatory analysis

To evaluate whether our results were solely dependent on the order of the steps, we performed a secondary, confirmatory analysis, designed so it would be “order agnostic.” We looked at the overlap in results from analyzing SNP data for the genes significant in step 2 versus four outcomes: (1) asthma; (2) intermediate phenotypes; (3) gene expression (eQTLs); and (4) cytokine levels. We additionally adjusted this analysis by WBC differential (percent of neutrophils, lymphocytes, monocytes, eosinophils, basophils). Rather than reducing the number of SNPs at each step, we looked at the overlap in results for the four outcomes. There were 148 SNPs from 40 genes (Figure 3 and Table E4) associated with at least three of these four outcomes. There was only one SNP, rs71058675 in IL5RA, associated with all four outcomes. There was also only one eQTL, rs1153462 (also in IL5RA), associated with asthma and intermediate phenotypes. Thus, this analysis confirms that our results are not entirely dependent on the selected order of the analytical steps. Of the other top genes, SOCS2 was associated with only two of the four outcomes, and IL1RN with one.

Figure 3.

Results from confirmatory analysis. Shown is the number of SNPs associated with each of four outcomes: asthma; intermediate phenotypes; cytokine levels; and gene expression. All results are from adjusted regression models adjusted for relevant covariates and principal components. Gene expression analysis additionally adjusted for white blood cell differential cell types. A total of 148 SNPs from 40 genes (see Table E4) were associated with at least three of four outcomes. Only one SNP, rs71058675, in IL-5 receptor α (IL5RA), was associated with all four outcomes. There was also only one expression quantitative trait loci, rs1153462 (also in IL5RA), associated with asthma and intermediate phenotypes. CellAdj, adjusted for CBC cell differential; eQTL, expression quantitative trait loci; GWAS, genome-wide association study; MedPhn, asthma intermediate phenotype.

Discussion

In the present study, we demonstrate a multiomics approach able to identify IL5RA as an asthma-related gene. To our knowledge, this is the first study of childhood asthma that integrates several analytical stages—based on the concept of pathogenic steps—to assess the significance of a gene or group of genes.

In our analysis, IL5RA showed significant associations in all analytical steps. IL5RA SNPs were associated with eosinophil counts, IL-17F (which in turn is associated with asthma and intermediate phenotypes), and asthma, as well as with IL5RA expression and IL-5Rα plasma levels. IL5RA methylation was associated with asthma. IL5RA expression was associated with eosinophils, STR, and its association with IL-17F was modified by asthma. In addition, IL-5Rα levels were associated with asthma age of onset and moderate–severe exacerbations. Of note, save for the association between IL5RA expression and eosinophils, most of these associations would have been missed in unidimensional analyses that customarily use statistical methods to adjust for multiple markers (e.g., for IL5RA SNPs versus asthma, the top SNP [rs17878498] had a value of P = 0.019). Although these steps are certainly not independent from each other, we propose that this “vertical validation” method, which evaluates significance at multiple pathogenic levels, is preferable to identify genes that are of greater biological plausibility and relevance. Furthermore, we observed associations between specific SNPs and several transcription factors, which show that these SNPs may have a direct regulatory role within known IL-5Rα signaling pathways.

The IL-5R, expressed on eosinophils, mast cells, and basophils, is a heterodimer composed of subunits IL-5Rα (specific to the IL-5R) and IL-5Rβ (common to the IL-3, IL-5, and granulocyte-macrophage colony–stimulating factor receptors) (29). We previously reported that an SNP adjacent to IL5RA is associated to the age of onset of asthma (30). The soluble form of IL-5Rα can occur by cleavage of existing cell surface receptors or by alternate splicing of IL5RA mRNA (31), and is a high-affinity competitive antagonist of IL-5 that prevents IL-5 from binding to surface IL-5R, and thus decreases activation and survival of eosinophils in vitro (32). Clinical studies have shown changes in soluble IL-5Rα levels in association with eczema in children (33), chronic obstructive pulmonary disease (34), and nasal polyposis (35). The balance between surface and soluble IL-5Rα has been shown to play a part in eosinophil regulatory pathways (31), and it could have implications for the use of anti–IL-5 therapeutic agents in asthma (36, 37). IL5RA was among the top results in a recent epigenome-wide association study of immune IgE (38). Anti–IL-5Rα monoclonal antibody, benralizumab, has been recently reported to reduce exacerbations and improve lung function in subjects with severe uncontrolled asthma (39, 40).

Our study has several limitations. Although we used the concept of disease pathogenesis to design our analytical approach, it is critical to underscore that our cross-sectional analysis evaluates associations and cannot adjudicate causality. GWE and GWM were not available for all study participants, and thus we had a relatively small sample size for some of the analyses, particularly those involving GWE, and thus limited statistical power to detect modest effects. We subjectively decided to use the gene expression–protein level as our first analytical step, hypothesizing that those two would be most closely related. However, it is important to clarify that we are not proposing that this is the only way or specific order in which these analyses must be undertaken; we simply propose that a “vertical validation” (e.g., significance at multiple pathogenic levels using multiomics integration) may render more relevant results than “horizontal replication” (e.g., attempting to replicate results from one analysis in a separate sample). This vertical validation approach does not expand the spectrum of findings across platforms, but rather uses these platforms to sequentially validate our findings. Pathway analysis may have biased our results toward better-characterized genes. We did not correct for multiple comparisons at all analytical steps; because the steps are not fully independent from each other, multiple testing correction at each level would have likely been overstringent. Future studies may be necessary to explore the optimal inferential approach to account for this problem. Finally, our analysis might have yielded different results had we proceeded in a different analytical order. However, we performed a secondary analysis that was agnostic to the order of the analytical steps, and IL5RA remained the top result. Furthermore, we propose that those results from analyzing the data in yet another order would likely be biologically relevant as well.

In summary, we present a multiomics approach that integrates several pathogenic steps and is able to identify biologically relevant asthma-related genes, such as IL5RA. Our approach differs from other methods for integrative analysis that rely on complicated statistical models with various assumptions; each step in our method uses linear or logistic regression, and is simple to interpret. Further studies will need to evaluate and validate this approach in diverse settings, including other complex diseases with multiomics data.