Gene Expression in Relation to Exhaled Nitric Oxide Identifies Novel Asthma Phenotypes with Unique Biomolecular Pathways

Abstract

Rationale

Although asthma is recognized as a heterogeneous disease associated with clinical phenotypes, the molecular basis of these phenotypes remains poorly understood. Although genomic studies have successfully broadened our understanding in diseases such as cancer, they have not been widely used in asthma studies.

Objectives

To link gene expression patterns to clinical asthma phenotypes.

Methods

We used a microarray platform to analyze bronchial airway epithelial cell gene expression in relation to the asthma biomarker fractional exhaled nitric oxide (Fe_NO) in 155 subjects with asthma and healthy control subjects from the Severe Asthma Research Program (SARP).

Measurements and Main Results

We first identified a diverse set of 549 genes whose expression correlated with Fe_NO. We used k-means to cluster the patient samples according to the expression of these genes, identifying five asthma clusters/phenotypes with distinct clinical, physiological, cellular, and gene transcription characteristics—termed “subject clusters” (SCs). To then investigate differences in gene expression between SCs, a total of 1,384 genes were identified that highly differentiated the SCs at an unadjusted P value < 10⁻⁶. Hierarchical clustering of these 1,384 genes identified nine gene clusters or “biclusters,” whose coexpression suggested biological characteristics unique to each SC. Although genes related to type 2 inflammation were present, novel pathways, including those related to neuronal function, WNT pathways, and actin cytoskeleton, were noted.

Conclusions

These findings show that bronchial epithelial cell gene expression, as related to the asthma biomarker Fe_NO, can identify distinct asthma phenotypes, while also suggesting the presence of underlying novel gene pathways relevant to these phenotypes.

At a Glance Commentary

Scientific Knowledge on the Subject

Global gene expression in asthma, especially in relation to clinical phenotypes, has not been fully explored.

What This Study Adds to the Field

In this study, we analyzed gene expression profiles of bronchial airway cells in relation to exhaled nitric oxide from patients with asthma and healthy control participants. This approach characterized clinically meaningful asthma phenotypes and identified novel molecular pathways for future investigation.

Severe asthma affects approximately 10% of the total asthma population but accounts for 50% of the disease-associated costs (1). The diagnosis of severe asthma, like asthma itself, is based on a broad definition that sheds little light on the underlying molecular pathobiology or phenotypic differences (2). Prior efforts to classify and understand asthma phenotypes have used mostly clinical phenotypic data (3–8), with little integration of more complex biologic information.

Gene expression microarrays, a snapshot of genome-scale steady-state messenger RNA (mRNA) expression, have been used to identify novel disease subtypes, predict drug response, and suggest pathological mechanisms in diseases such as cancer and inflammatory disorders (9, 10). In asthma, expression of three bronchial airway epithelial cell (BAEC) genes was used to identify an asthma phenotype (termed “Th2 high”) among a group of corticosteroid (CS)-naive patients with mild asthma associated with elevated serum IgE, CS responsiveness, and increased blood and airway eosinophils (11, 12). Molecular phenotyping beyond these three epithelial genes in a broader range of patients with asthma has not yet been performed.

Fractional exhaled nitric oxide (Fe_NO), an asthma biomarker that associates with atopy and eosinophilia (6, 13), is strongly induced by type 2 cytokines in vitro. Its inhibition in vivo by biologic therapies targeted toward IL-4 and -13 confirms its association with this immune pathway (14–17). However, Fe_NO elevation is observed across a range of disease states, including a high-risk asthma phenotype, CS refractory asthma, and even atopy without asthma (18, 19). These findings limit its clinical usefulness (20–22) and suggest that multiple biological factors beyond type 2 inflammation may regulate its production and/or downstream consequences.

We hypothesized that clinically distinct asthmatic phenotypes could be identified within a diverse population of subjects by clustering BAEC samples according to the expression of a subset of genes correlated with Fe_NO. We further hypothesized that clustering of genes that are most differentially expressed between these phenotypes would provide biologic information on disease mechanisms. To address this hypothesis, bronchoscopically obtained epithelial brushings from extensively characterized subjects with asthma and healthy subjects in the National Heart Lung and Blood Institute–sponsored Severe Asthma Research Program (SARP) were analyzed by gene expression microarray. Genes strongly correlating with Fe_NO levels were used to cluster epithelial samples and identify potentially meaningful clinical subject clusters (SCs). Additional genes not correlating with Fe_NO, but still differentiating the SCs, were then clustered to explore molecular and biologic pathways specific to each SC. Some of the results have been previously reported as an abstract (23).

Methods

See online supplement for complete details.

Study Population

Samples and data were obtained and characterized from subjects in SARP from 2009 to 2011. In this cross-sectional observational study, adherence to prescribed medications was not formally addressed in the period before sampling. However, coordinators and investigators informally addressed compliance to inhaled and systemic CSs over the course of three to four study visits. Subjects believed to have poor adherence to medications were generally excluded.

Oligonucleotide Microarray Experiments

The complete data set is available online (National Center for Biotechnology Information’s Gene Expression Omnibus database; http://www.ncbi.nlm.nih.gov/geo/; accession numbers GSE63142 and GSE43696). Low-variance probes were removed and median values used for probes with identical RefSeq numbers, resulting in a filtered list of 19,567 genes.

Real-Time Quantitative Reverse Transcription–Polymerase Chain Reaction

Real-time quantitative reverse transcription–polymerase chain reaction (qRT-PCR) confirmed microarray gene expression with RNA extracted from the same bronchial brushings. Genes validated included those identified in the clustering analysis (n = 19) as well as additional genes of potential relevance to asthma (n = 5).

Subject Sample Clustering by Fe_NO-correlated Genes

Fe_NO level was not normally distributed and therefore correlated with Fe_NO using Spearman correlations at a threshold false discovery rate of 5%. k was chosen by a previously described method that minimizes the difference between error sum of squares for the array data and a randomized set of data (24). As shown in Figure E1, the rate of improvement (slope of the curve) tapers as k reaches the range of 4 to 7. The variable k chosen (n = 5) was the largest value that resulted in at least 10 subjects per cluster, a minimum size needed to statistically detect phenotypic differences.

Comparison of SC Phenotype Data after k-Means Clustering

Phenotypic differences between SCs were detected by statistical methods in the online data supplement. All significant intergroup comparisons listed in the tables achieve the appropriate Bonferroni corrected P value of 0.005 or less (for comparisons among five groups).

Identification of Genes Most Differentially Expressed between SCs

After establishment of SCs, all genes expressed in the epithelial brushings (total = 19,957) were compared across SCs for differential expression using intergroup Student t tests. Genes differentially expressed to a P value < 10⁻⁶ between at least two of the SCs were used to compose a set of “highly differentiating” genes. Of note, 149 of the original 549 Fe_NO-correlated genes were not included, as they did not meet this criterion. To ensure that the genes in the highly differentiating gene set were not identified by “chance,” epithelial samples were randomly reassigned to same-sized SCs with gene expression comparisons repeated for five iterations. In these randomly reassigned SCs, there were no genes differentially expressed at a P value < 10⁻⁶ between any two of the five SCs.

Hierarchical Clustering of Genes to Explore Gene Clusters

As mentioned, a highly differentiating gene list was generated by finding the most differentially expressed genes across the five SCs. Hierarchical clustering using only these highly differentiating genes was then performed to determine genes with similar expression patterns (i.e., “coexpressed” genes). Based on the hierarchical clustering results, these genes were placed into gene clusters (GCs), and analyzed using Ingenuity and PubMed gene annotation.

Results

Demographics by Traditional Asthma Severity Criteria

Samples and data from 155 participants were obtained from 2009 to 2011 (Table 1). There were 26 healthy control subjects, and 51 participants met the American Thoracic Society 2000 definition of severe asthma. There were no differences in race and sex across the groups. Subjects with severe asthma (SA) were older than subjects with mild to moderate asthma/no inhaled CSs (ICS) and healthy control subjects (HCs), and had a higher body mass index (BMI) than the HCs. FEV₁ decreased with severity. Although blood IgE and Fe_NO differed among the groups, there were no intergroup differences. Of the subjects with SA, 63% used oral CSs (OCS) on a regular basis, supporting the severity of disease in these patients.

Table 1.

Demographics by Traditional Clinical Characteristics

	HCs	Mild-Mod No ICS	Mild + ICS	Mod + ICS	Severe	P Value	Significant Intergroup Comparison*
Age, mean ± SD	32 ± 12	29 ± 10	37 ± 12	36 ± 12	45 ± 11	<0.0001	SA > HC
							SA > Mild no ICS
Sex, M/F	12/14	13/24	6/16	8/11	16/35	0.61	No intergroup difference
Race, W/AA/O	18/5/3	21/10/5	14/5/2	8/10/1	31/14/3	0.43	No intergroup difference
BMI, median (25–75th quartile)	24 (22–30)	28 (23–32)	28 (25–34)	29 (25–32)	32 (26–37)	0.006	SA > HC
FEV1% predicted, mean ± SEM	94 ± 9	84 ± 16	91 ± 16	68 ± 8	56 ± 20	<0.0001	HC > SA
							Mild no ICS > SA
							Mild + ICS > SA
							HC > Mod + ICS
							Mild no ICS > Mod + ICS
							Mild + ICS > Mod + ICS
Atopy, yes/no	16/10	31/4	15/4	16/2	39/8	0.08	No intergroup difference
IgE, median (25–75th quartile), IU/ml	59 (11–132)	113 (54–432)	139 (56–468)	142 (55–314)	104 (26–434)	0.042	No intergroup difference
Blood eos, median (25–75th quartile), %	2 (2–3)	3 (2–5)	2 (2–5)	4 (2–7)	3 (1–5)	0.14	No intergroup difference
FeNO, median (25–75th quartile), ppb	20 (13–37)	37 (19–60)	21 (13–39)	39 (22–59)	37 (21–63)	0.043	No intergroup difference

Definition of abbreviations: AA = African American; BMI = body mass index; eos = eosinophils; F = female; Fe_NO = fractional exhaled nitric oxide; HC = healthy control subject; ICS = inhaled corticosteroids; M = male; Mod = moderate; O = other; SA = patient with severe asthma; W = white.

^*All significant intergroup differences listed achieved a Bonferroni corrected P value of P ≤ 0.005.

Identification of Genes Correlating with Fe_NO

One hundred twenty-seven of 155 subjects had Fe_NO data available. Correlations of Fe_NO with all differentially expressed genes (n = 19,567) identified 549 genes (see Table E2 in the online supplement) that correlated with Fe_NO (false discovery rate < 0.05), ranging from 0.3 to 0.72 for positive correlations and −0.54 to −0.3 for negative correlations. The enzyme most likely responsible for NO production, inducible nitric oxide synthase (NOS2), had the strongest correlation (rho = 0.72, P value < 10⁻⁷).

Identification of SCs

Using the subset of genes that correlated with Fe_NO, samples were clustered by k-means with k = 5 (Figure 1). This choice was supported by the markedly different characteristics of the resulting five SCs shown in Table 2 and Figure 2. Three SCs had abnormally high Fe_NO levels (SC2, 3, and 5), and two had normal Fe_NO levels (SC1 and 4). Nearly 90% of all HCs are in SC1 and 4, and 65% of traditionally defined subjects with severe asthma are in SC2 and 3 (Figure 2).

Figure 1.

Heat map derived by k-means clustering of subject samples with genes that correlated with exhaled nitric oxide (Fe_NO). Gene expression was correlated with Fe_NO using a Spearman correlation method, identifying 549 genes that correlated strongly (positively and negatively) with Fe_NO. k-means clustering of subject bronchial epithelial brushing samples with these genes identified five unique SCs shown here (x axis, top).

Table 2.

Demographics by Subject Clusters

	SC1	SC2	SC3	SC4	SC5	P Value	Significant Intergroup Comparisons*
FeNO, median (25–75th quartile), ppb	18 (12–25)	55 (38–103)	39 (18–50)	19 (14–27)	38 (29–75)	<0.0001	SC2 > SC1
							SC3 > SC1
							SC5 > SC1
							SC2 > SC3
							SC2 > SC4
Age, mean ± SD, yr	36 ± 12	38 ± 13	42 ± 14	42 ± 12	27 ± 9	0.0013	SC1 > SC5
							SC2 > SC5
							SC3 > SC5
							SC4 > SC5
Sex, M/F	15/41	18/29	11/12	5/7	6/11	0.44	No intergroup difference
Race, W/AA/O	36/15/3	25/15/6	14/4/3	9/1/2	8/9/0	0.15	No intergroup difference
BMI, median (25–75th quartile)	28 (22–32)	31 (27–37)	30 (24–37)	31 (25–34)	25 (21–29)	0.0024	SC2 > SC1
							SC2 > SC5
FEV1% predicted, mean ± SEM	85 ± 19	66 ± 21	66 ± 22	76 ± 26	83 ± 16	<0.0001	SC1 > SC2
							SC1 > SC3
							SC5 > SC2†
Change after BD, median (25–75th quartile), %	7 (4–14)	19 (10–41)	15 (11–21)	10 (5–14)	11(6–16)	0.0001	SC1 > SC2
							SC1 > SC3
Atopic sensitization, yes/no	34/18	37/7	18/2	12/0	16/1	0.0070	No intergroup difference
Number of positive skin tests, mean ± SD	3 ± 3	4 ± 3	4 ± 3	5 ± 3	5 ± 1	0.016	No intergroup difference
IgE, median (25–75th quartile), IU/ml	49 (15–106)	142 (51–433)	232 (102–342)	66 (40–266)	294 (169–783)	<0.0001	SC2 > SC1
							SC3 > SC1
							SC5 > SC1
Blood eos, median (25–75th quartile), %	2 (1–3)	4 (3–6)	2 (1–5)	3 (2–3)	4 (3–8)	<0.0001	SC2 > SC1
							SC5 > SC1
							SC5 > SC4‡
Uses oral CS, %	4	27	47	20	14	0.0019	No intergroup difference

Definition of abbreviations: AA = African American; BD = bronchodilator; BMI = body mass index; CS = corticosteroids; eos = eosinophils; F = female; Fe_NO = fractional exhaled nitric oxide; M = male; O = other; W = white.

k-means clustering of subjects with genes (n = 549) that correlated strongly (positively and negatively) with exhaled nitric oxide (Fe_NO) identified five unique subject clusters (SCs). This table details the demographics of these newly identified SCs.

^*All significant intergroup differences listed achieved a Bonferroni corrected P value of P ≤ 0.005.

^†P = 0.0074.

^‡P = 0.0059.

Figure 2.

Composition of subject clusters (SC1–5) according to traditional asthma severity criteria. Gene expression was correlated with exhaled nitric oxide (Fe_NO) using a Spearman correlation method, identifying 549 genes that correlated strongly (positively and negatively) with Fe_NO. k-means clustering of subject bronchial epithelial brushing samples with these genes identified five unique SCs. As shown in the figure, SCs greatly differed according to traditional asthma severity criteria. SC2 and 3 had the highest percentage of patients with severe asthma. HCs = healthy control subjects; ICS = inhaled corticosteroids; Mild-Mod no ICS = mild to moderate asthma not on ICS; Mild + ICS = mild asthma on ICS; Mod + ICS = moderate asthma on ICS; Severe = severe asthma.

Description of SCs

See Tables E3–E5 for differences in specific characteristics.

SC1.

SC1, the largest cluster, contains 73% of all HCs in the study, and more than two-thirds are either HCs or subjects with mild asthma on no or low-dose ICS. Fifty percent of all subjects with mild asthma treated with ICS are in this cluster, suggesting that effective ICS therapy may nearly normalize epithelial gene expression. Despite this, 16% of subjects in this cluster are traditionally classified as having SA. Consistent with the large percentage of HCs in this cluster, lung function is near normal (FEV₁ = 85 ± 16%), and asthma symptoms are infrequent. There is little evidence for inflammation according to Fe_NO (median = 18 ppb), blood, or bronchoalveolar lavage (BAL) cells. Atopy and IgE levels (median = 49 IU/ml) are also relatively low (Table 2).

SC2.

SC2 is the second-largest SC, with the highest Fe_NO overall (median = 55 ppb). Traditionally defined moderate to severe asthma makes up 73% of this SC, with low lung function (FEV₁ = 66 ± 21%), a high degree of symptoms, the greatest reversibility (19%; 25–75th percentile, 10–41%) and a high percentage on regular OCS (27%). Eosinophils in blood and BAL were higher than those in SC1 and 4. Despite the highest Fe_NO and high blood eosinophil percentages, IgE level (median = 142 IU/ml) and number of positive skin tests are intermediate among the SCs. Subjects in this cluster generally developed disease in childhood (median, 8.5 yr; 25–75th percentile, 4–18 yr) and are obese.

SC3.

The third-largest SC is arguably the most severe. Seventy-four percent have moderate to severe disease by traditional measures, with 56% classified as severe. Similar to SC2 above, SC3 has low lung function (FEV₁ = 66 ± 22%) and high Fe_NO (median = 39 ppb) but tends to have a later age at onset (13 yr; 25–75th percentile, 2–31 yr) and shorter duration of disease. There is little history of asthma in a parent (10% of subjects), although IgE levels are high (median = 232 IU/ml). Despite containing the highest percentage of patients receiving regular OCS (47%), BAL cell eosinophilia is similar to the other Fe_NO-high clusters. Blood eosinophilia tends to be lower than the other Fe_NO-high clusters (2%), consistent with a systemic CS effect. SC3 has the highest levels of peripheral blood and BAL neutrophils (62 and 4.9%, respectively), also consistent with high OCS use. It also has the second highest percent of BAL lymphocytes, which suggests a mixed inflammatory milieu in the lungs. SC3 has the highest percent with a history of nasal polyps (26%) and sinus surgery (30%) (Table E5) and is similar to SC2 in BMI.

SC4.

SC4 has normal Fe_NO levels (median = 19 ppb) and is the smallest SC (n = 12). Although 33% are HCs (n = 4), in comparison to cluster SC1, which contains 73% of all HCs, this represents only 15% of all HCs. Unlike SC1, this cluster consists of 50% patients with moderate to severe asthma, with associated high health care use (27% ever intubated) and moderately high systemic CS use (20%). After removing the HCs, the lung function is low (FEV₁ = 66 ± 7%) and associated with the earliest age at onset (3 yr; 25–75th percentile, 0–6 yr) and the longest duration of disease (median = 34 yr; 25–75th percentile, 33–54 yr). Notwithstanding low Fe_NO, blood and BAL eosinophils, and total IgE (median = 66 IU/ml), 100% are atopic, and the number of positive skin tests does not differ from the high Fe_NO clusters. Despite the lack of eosinophilic or neutrophilic inflammation, this cluster has the highest BAL lymphocytes (median = 13.2%). The median BMI is the highest of any SC but does not differ from other SCs, likely due to its small sample size.

SC5.

SC5 is the youngest group (mean age = 27 yr), with an elevated Fe_NO (38 ppb), having an early age at onset (6 yr; 25–75th percentile, 2–11 yr) and shorter duration of disease than in SC4. More than 50% identify as African Americans (highest in any cluster), with the highest percentage of any cluster reporting asthma in a parent (59%). Only 24% have severe disease by traditional measures, and they have normal lung function (FEV₁ = 83 ± 16%), similar to SC1. Despite this, 35% report an ICU stay or intubation. As would be consistent with their younger age and African racial background, IgE levels are highest in this group (median = 294 IU/ml), and eosinophilic inflammation in both blood and BAL is relatively high and similar to SC2. There is very little OCS use (14%) or history of nasal polyps/sinus disease. Perhaps reflective of their younger age, this SC is the leanest of all SCs (median BMI = 27).

GC Results

Complete linkage hierarchical clustering of the highly differentiating set of genes (n = 1,384) along the y axis of the heat map identified nine clusters of coexpressed genes using a cutline at the fourth division of the hierarchical tree (Figure 3). The biologic characteristics of the clusters are summarized below and in Table 3. A complete listing by GC is found in Tables E7.1–7.9. Twenty-one genes were validated by mulitplex qRT-PCR, although not all of them were used in this study. The majority (n = 17) demonstrated a strong correlation (Spearman correlation, r > 0.5) (Table E6).

Figure 3.

Complete linkage hierarchical clustering of genes that most strongly differentiate the subject clusters (SCs). After establishment of SCs, all genes expressed in the epithelial brushings (total = 19,957) were compared across SCs for differential expression using intergroup Student t tests. Genes differentially expressed to a P value < 10⁻⁶ between at least two of the SCs were used to compose a set of “highly differentiating” genes (n = 1,384) shown in the heat map here. Complete linkage hierarchical clustering of this highly differentiating gene set along the y axis of the heat map identified nine clusters of coexpressed genes using a cutline at the fourth division of the hierarchical tree (y axis, left). SCs remain on the x axis. GC = gene cluster.

Table 3.

Gene Cluster Molecular and Biological Enrichment Characteristics

GC (n)	High (SCs)	Int (SCs)	Low (SCs)	Functional Pathway/Cellular Association	Genes
1 (131)	1	3, 4, 5	2	Innate immunity/antibacterial function	BPIFA1, MUC5B, PGLYPR4, NOX1
				Cell proliferation/apoptosis	ERBB2, S1PR, IL36RN
				Lymphocyte activation/migration	LAG3, CCL15, SLAMF7, SLAMF9, MALL, TNFSF4
				Other	HCRT, FGF family
2 (172)	4	1, 5	2, 3	Cilia structure/function	DNAH1, DNAH3, IFT122, DNAI1, FBF1, TMEM231, BBS1, ARL3, IFT140, HYDIN, KIF19, DNAH10
				Other	CRY2, INSR, SCGB1A1 and 3A1, SMAD3, DTX3, TMEM8B, ATXN2
3 (52)	4, 3	2	1, 5	TNF-α	SPPL2B, MAPK4K4, IKBKB
				Muscle	DMPK. POPDC2, MUSTN1, TNNI3, FERL5, MYO15B
4 (325)	1	4, 5	2, 3	Notch signaling/cell adhesion	WNK4, JAG2, EPDR1, CD302, CDON, GXYLT2, HEY2, THBS4, RBPJ, CLDN16, OSBPL3, GJB7, CDH1
				Neuronal function	SHANK2, LRP8, KCNN3, MAOB, SEMA3E, LMO3, EFHD1, NDN, CHN1, NPAS3, NDNF, NAV3, SLITRK5, SLC6A16, SNCAIP, TUB, INA, MAPT, ORS1E1, SCSK1, EFNB2, PMMA3, CLU, STXBP1, ENO2, OPRL1, DDC
				Dystrophin family	SNTB1, SGCE, SSPN, CNTD1
				WNT family	WIF1, WNT5B, RHOU, ASCDD1, DKK3
				Ion channels	KCNA1, KCNC1, CACNB1, CACNA202, SCN3A, NEDD4L
				Other	BPHLHE31, PER2, IL5RA, GSTA1, 2, 3, 5, HNMT, GAB2, PIAS1, FOXA2, LTF
5 (50)	1, 5	2, 4	3	No obvious functional/cellular associations “Interesting genes”	PFDN5, KCNMB2, BTK, TNFSF12, GPAN
6 (150)	2, 3	4,5	1	Microtubules (2%)	TUBA4A, ACOT13, CKAP2
				Mitochondrial (3%)	ATP5H, MTERF, ETFA, NNT, CPT1A. CYCS
				Actin related (2%)	LIMA1, PDF1B, ACTR3, TMP4
				Neuronal (2%)	BDNF, GRIK2, NAV1, RIT1
				Other	IRAK4, GART, GMP5, SMAP1, ADAM9, IBTK, ITGB6, CXCL14
7 (228)	2, 3	4	1, 5	Interferons (type 1 and 2) (∼2%)	DUOX2, MX2, IL18R1, NOS2
				Apoptosis (∼3%)	CARD14, TRADD, CASP6, TXNDC5, AK2, TNFRFS18, MAGED1
				P38 related	MAPK13, MAP2K6
				Keratins	KRT23, KRT6A, KRT18
				Sialyl Lewis antigen	FUT3, 5, 6
				Cell matrix interactions (∼3.5%)	DIAPH3, LGALS7, LGALS7B, FERMT1, LOXL4, RHOC, LOX, TNC
				Other	IRAK3, TEP and TIMELESS, PTK6, MUC4, DEFB1, PLA2G4A and ABCC1, LRRC8A, ICA1, POP1
8 (96)	2, 5		1, 3, 4	Cysteine metabolism (4%)	CCBL1, CST1, CST2, FETUB
				Mucins (5%)	MUC5AC, TFF3, TFF2, FOXA3, GANLT12
				Mast cells (11%)	HPGDS, PTGS1, P2RY14, CTSG, HDC, IL1RL1, KIT, CPA3, TPSAB1, MS4A2, TPSD1, CD22/SIGLEC2, SIGLEC17P, SLC18A2, DPP4
				Vasoconstrictors (possibly MC) (3%)	UTS2, P3RX1
				Glycolipid antigen presentation	CD1C, CD1E
				Other	PTGDR2, MYO1C, CNRIP1, PRB1, 2 and 4
9 (180)	2	3, 4, 5	1	Mitochondria (3%)	MTHFD2, MRPL15, COX17, ATP5O, PYCR1, HK2
				Intracellular trafficking (5%)	STKBP5, RAB27B, TBC1D15, S100A6. ARL1, KIF16B, SYTL4, SEC24D, SEC22A
				O-linked glycosylation (2%)	GALNT1, 4, 7, 10
				N-linked glycosylation (2%)	MAGT1, FUT8, ALG5, ERLEC1
				“Type 2 genes” (2%)	ALOX15A, CCL26, SERPINB2, POSTN
				Other	BCL2L14, RIPK2, CORO1C, ARNTL2, MUC2, MGST2, NTRK2, CD44

Definition of abbreviations: GC = gene cluster; Int = intermediate; SCs = subject clusters; TNF = tumor necrosis factor.

Genes differentially expressed to a P value of P < 10⁻⁶ between at least two of the five SCs were added to the Fe_NO-correlated genes to form a highly differentiating set of genes. Complete linkage hierarchical clustering of this gene set formed nine GCs. This table details important genes with associated functional pathways and/or cellular associations in each of these GCs. The columns “High,” “Int,” and “Low” indicate comparative expression levels.

GC1 (n = 131) (High = SC1; Intermediate = SC3, 4, and 5; Low = SC2)

GC1 includes genes of innate/antimicrobial defense and adaptive immunity (3% combined), including: BPI fold containing family A, member 1 (BPIFA1) (validated by qPCR) (25) and mucin 5B (MUC5B) (26–28). GC1 is also enriched for genes related to cell growth, proliferation, and survival (2%), including the epidermal growth factor receptor gene ERBB2 (validated by qPCR) and sphingosine-1-phosphate receptor (S1PR). The #1 and #2 ranked Ingenuity canonical pathway genes involve immunologic signaling and leukocyte mobility (Table E8).

GC2 (n = 172) (High = SC4; Intermediate = SC1 and 5; Low = SC2 and 3)

GC2 is highly enriched for cilia-related genes (12%), including dynein motor proteins and subunits of the intraflagellar transport complex, a complex responsible for the genesis, resorption, and signaling of primary cilia.

GC3 (n = 52) (High = SC4 and 3; Intermediate = SC2; Low = SC1 and 5)

This small GC is highly enriched for muscle-related genes (12%) and tumor necrosis factor (TNF)-related signaling (6%).

GC4 (n = 325) (High = SC1; Intermediate = SC4 and 5; Low = SC2 and 3)

This large GC4 is enriched with diverse functions and pathways. Surprisingly, 8% of genes associate with neuronal development and function. GC4 also includes NOTCH (4%) and WNT developmental pathway genes. The #5 ranked Ingenuity upstream analysis is WNT signaling protein (WNT3A, P = 4.89 × 10⁻⁴; Table E8). Finally, GC4 contains several reducing enzymes, and the #1 ranked Ingenuity pathway is glutathione-mediated detoxification (P = 7.8 × 10⁻⁴), including Glutathione S-transferase α 2 (GSTA2), validated by qRT-PCR.

GC5 (n = 50) (High = SC1 and 5; Intermediate = SC2 and 4; Low = SC3)

Although some well-described genes are in this subset, including Bruton’s tyrosine kinase (BTK), there is not enrichment for a known molecular or cellular process. Likewise, there was no significant pathway identified in Ingenuity.

GC6 (n = 150) (High = SC2 and 3; Intermediate = SC4 and 5; Low = 1)

GC6 is enriched with mitochondria- (3%), microtubules- (2%), and actin- (2%) related genes. Affirmatively, the #3, #4, and #6 ranked Ingenuity pathways include oxidative phosphorylation (4.2 × 10⁻²), mitochondrial dysfunction (2.8 × 10⁻²), and an actin-based pathway (4.5 × 10⁻²), respectively. This cluster also contains neuronal genes (2%), including brain-derived neurotrophic factor (BDNF). The #1 ranked Ingenuity upstream analysis is neuronal development gene atrophin 1 (ATN1) (2.9 × 10⁻⁵).

GC7 (n = 228) (High = SC2 and 3; Intermediate = SC4; Low = SC1 and 5)

GC7 includes inducible NOS2, the gene correlating most strongly with Fe_NO. Affirmatively, the #3 Ingenuity pathway is iNOS signaling (P = 1.12 × 10⁻³). GC7 is also enriched for type 2 and 3 interferon genes (2%), including IL-18RA and DUOX2 (dual oxidase-2) genes, validated by qRT-PCR. Ingenuity upstream analysis supports a prominent effect from TNF (P = 9.62 × 10⁻⁶), likely contributing to the enhancement of genes associated with apoptosis, including TRADD (TNFRSF1A-associated via death domain) and CARD14 (Caspase Recruitment Domain Family, Member 14). At least 3% of genes in GC7 are associated with cell–matrix interactions.

GC8 (n = 96) (High = SC2 and 5; Low = SC1, 3, and 4)

GC8 is enriched for mast cell–associated genes (11%), including TPSAB1 (tryptase) (tryptase genes validated by qRT-PCR), MS4A2 (β chain of high-affinity IgE receptor), and HPGDS (hematopoietic prostaglandin D2 [PGD2] synthase), with their tight clustering demonstrating strong evidence for coexpression. CRTH2, one of the receptors for PGD2, is also found here. Other genes include mucin-related (including MUC5AC, validated by qRT-PCR) (5%), cysteine metabolism (4%), vasoconstriction, and glycolipid antigen presentation (CD1 genes). The #1 Ingenuity upstream analysis is cytokine IL-13 (P = 3.35 × 10⁻⁷).

GC9 (n = 180) (High = SC2; Intermediate = SC3, 4, and 5; Low = SC1)

GC9 contains the “Th2-like” genes periostin (POSTN), chemokine ligand 26 (CCL26), and 15 lipoxygenase 1 (ALOX15A). Both POSTN and CCL26 were validated by qRT-PCR. GC9 is also enriched for genes related to intracellular trafficking (5%), mitochondrial function (3%) and O-linked glycosylation (2%). The #1 ranked Ingenuity upstream analysis protein is transcription factor XBP1 (5.4 × 10⁻¹¹).

Discussion

This report details the first large-scale study to combine exhaled nitric oxide and bronchial epithelial cell gene expression to objectively cluster a broad range of patients with asthma into distinct phenotypes. Detailed clinical, physiological, cellular, and gene transcription characteristics were combined to fully describe these newly defined clusters. This study is also unique in the use of gene expression patterns to uncover known and novel molecular pathways relevant to these asthma phenotypes.

Fe_NO has emerged as a viable asthma biomarker. NOS2 is the gene most strongly correlated with Fe_NO (rho = 0.72, P value < 10⁻⁷), supporting its known link to BAEC expression and the biologic plausibility of its correlation with BAEC gene expression profiles (Table E2) (29–31). Similarly, the “type 2” immunity–associated genes POSTN, CLCA1, and SERPINB2 also strongly correlate with Fe_NO (rho values from 0.49–0.60, P values all < 10⁻⁷), whereas the mucin gene MUC5B, which is strongly repressed by IL-13 (12), has the strongest negative correlation (rho = 0.54, P value < 10⁻⁷). This study confirms an association of Fe_NO with type 2 immune processes. However, it adds many genes negatively and positively correlated with Fe_NO, providing a diverse set of genes to identify SCs in subsequent cluster analysis.

k-means clustering of the 549 Fe_NO-correlated genes identified five molecularly defined and clinically distinct SCs, three with higher Fe_NO (SC2, 3, and 5) and the remaining two with low/normal Fe_NO levels (SC1 and 4) (Table 2; Tables E2–E5). The majority of HCs, including all of those without atopy, clustered into SC1. It is therefore considered to be a “healthy” cluster. The patients with asthma who clustered into SC1 primarily had mild or well-controlled disease. Traditionally defined severe asthma was most prevalent in SC2 and 3, identifying these two clusters as the most “severe” SCs. Both of these clusters have high Fe_NO levels, high inhaled and systemic CS use, BAL eosinophilia, and low lung function. However, when compared with SC2, SC3 has a later age at onset and more nasal polyposis and neutrophilic inflammation. SC4 is 100% atopic, has the earliest age of asthma onset, longest disease duration, and severe disease, but normal Fe_NO and a high percentage of BAL lymphocytes. SC5, the youngest cluster, has high Fe_NO, a high percentage of African Americans (>50%), and, perhaps not surprisingly, the highest IgE levels (32–34). Although eosinophil levels are high, lung function does not differ from the “healthy” SC1.

Hierarchical clustering with the highly differentiating gene set of 1,384 genes (Figure 3) identified nine GCs that helped deepen our molecular understanding of each SC. Exploration of GC composition helped identify transcriptional patterns in severe asthma SCs that indicated both ongoing epithelial injury as well as inflammatory processes. GC1, 2, and 4 are all expressed at lower levels in “severe” SC2 and 3 compared with SC1. GC1 contains critical host defense genes, such as BPIFA1 (25) and the mucin gene MUC5B (26–28), which likely translates to an increased susceptibility to infection among patients with more severe asthma. MUC5B deficiency is associated with chronic infections in mice (28), with low levels previously described in asthmatic airways in relation to type 2 inflammation (12). Similarly, suppression of cilia-associated genes (i.e., dynein motor proteins of GC2), cell adhesion, and wound repair–related genes (GC4) suggest abnormal airway epithelial defense/structure among those with more severe disease. Fascinatingly, 8% of GC4 genes relate to neuronal function, suggesting neural features are an integral part of a healthy epithelial layer, whereas loss of expression could contribute to disease pathogenesis.

GC transcription patterns also revealed inflammatory conditions unique to each SC. Phenotypically, there are more type 2 inflammatory characteristics (i.e., elevated blood eosinophils and Fe_NO [Table 2] and BAL eosinophils [Table E4]) in SC2 and 5 than in the “healthy” SC1. Supportively, they have increased expression of the mast cell–related GC8, and the “Th2-high” asthma phenotype genes found in GC9. Notwithstanding the fact that SC3 has an elevated Fe_NO and SC4 has high atopy (100%), both have relatively lower expression of the mast cell–associated GC8. This suggests that a mast cell signature is quite specific for atopic/allergic and inflamed asthma.

Genes in GC7 are more highly expressed in “severe” SC2 and 3 compared with “healthy” SC1. GC7 contains iNOS and related pathway genes. These include IFN-γ signaling, supporting prior reports that indicate higher levels of IFN-γ are found in “severe” asthma and suggesting an involvement of type 1 pathways in iNOS expression (35–37). GC7 also associates with TNF signaling. The enhancement of TNF-α pathways and apoptosis-related genes in GC7 support overall derangements of cell death–related properties.

Finally, GC3 is higher in “severe” SC3 and “moderate-to-severe” SC4 compared with “healthy” SC1. GC3 contains muscle-related genes and those related to TNF signaling. SC4 also has intermediate expression of GC7, which, as mentioned above, also contains genes related to TNF signaling. This supports the role of TNF-induced inflammation in SC3 and suggests it may be a predominant or novel disease mechanism in SC4. Whether these patients might selectively benefit from anti-TNF therapeutic approaches remains to be determined.

Limitations to the study include the cross-sectional nature and the lack of studies that associate this mRNA expression to protein expression. In addition, Fe_NO levels are influenced by post-transcriptional/translational modifications of iNOS and other relevant proteins, as well as by substrate levels. However, iNOS (NOS2) mRNA correlates to Fe_NO with a rho = 0.72, suggesting ∼50% of the Fe_NO level is explained transcriptionally. Cells were not sorted before RNA extraction, making it unclear whether mRNA expression was derived from epithelial cells or other cell types. However, in many cases, mRNA specific for certain cell types, particularly mast cells, goblet cells, and ciliated epithelial cells, clustered together, supporting their expression from a common cell type. Finally, both use of and nonadherence to CS therapy could impact the relevance of the transcriptional data reported here. Notwithstanding, despite having the highest reported CS use and other indications of systemic CS adherence (e.g., blood and BAL neutrophilia in SC3), SC2 and 3 demonstrated strong inflammatory signals, including high gene expression of iNOS and associated genes in GC7. Also, OCS (or ICS) use was not specific to any SC. Although the study is likely biologically pertinent, its clinical importance must be tested by the ability of the gene and SCs to predict response to therapy.

In conclusion, this was the first large-scale study to combine exhaled nitric oxide and bronchial epithelial cell gene expression to objectively identify novel asthma phenotypes while also uncovering unique and complex relationships between gene expression and Fe_NO. This analysis may help to (1) improve current understanding of underlying pathobiology among asthma phenotypes, and (2) lead to the development of biomarkers that enhance the usefulness of Fe_NO measurements. This hypothesis-generating study opens the door to future studies to confirm and/or refine these asthma phenotypes as well as identify novel molecular pathways in disease.