Molecular Phenotyping Based on Blood Transcriptome-Associated Clusters in Asthma: Findings From Two Korean Asthma Cohorts

Abstract

Purpose

Asthma is heterogeneous in its molecular and clinical manifestations, complicating treatment. In this study, we investigated the molecular phenotypes of asthma using RNA-sequencing (RNA-seq) of peripheral blood mononuclear cells to advance precision medicine approaches.

Methods

We analyzed peripheral blood RNA-seq data from 456 participants from the Cohort for Reality and Evolution of Adult Asthma in Korea, including 23 patients with severe asthma and 433 with non-severe asthma to identify blood transcriptome-associated clusters (B-TACs) that correlate with distinct inflammatory processes. This included a meta-analysis of differentially expressed genes to refine gene selection for defining B-TACs. Gene Set Variation Analysis was used to assess the differential activation of biological pathways across B-TACs. We validated these findings in the Precision Medicine Intervention in Severe Asthma cohort of 122 patients, focusing on cluster stability, clinical correlates, and treatment responsiveness.

Results

Our analysis identified 3 B-TACs (B-TAC1, B-TAC2 and B-TAC3) with B-TAC1 and B-TAC2 showing higher blood eosinophil levels than B-TAC3. B-TAC1 was characterized by a clear T2-high inflammatory profile, whereas B-TAC2 exhibited a mixed molecular phenotype with both T2-high and non-T2 features. B-TAC3 was associated with low T2 and inflammasome-related inflammation profiles. Validation using the PRISM cohort underscored the stability of B-TAC1 and B-TAC2. Notably, patients with B-TAC2 molecular phenotype had lower biologics use and tended to have less favorable treatment outcomes than those with B-TAC1.

Conclusions

We discovered and validated 3 B-TACs, with B-TAC2 displaying a mixed T2-high and non-T2 phenotype, suggesting a potential reduction in responsiveness to biologic treatments. These findings provide molecular insights that may guide the development of personalized treatment strategies for asthma.

INTRODUCTION

Asthma is a heterogeneous disease influenced by genetic predispositions and environmental factors. The inflammatory mechanisms in asthma are highly diverse, involving numerous pathways and multiple cellular components intricately interacting to drive the disease.¹ Traditional uniform treatments typically fail to provide sufficient therapeutic effects owing to diverse responses among patients.² Chronic inflammation, which triggers asthma, manifests through multiple biological pathways, with well-known mechanisms such as T2 inflammation.³ For instance, eosinophilic inflammation often results from allergic responses or innate inflammation, driven by Th2 cells and type 2 innate lymphoid cells (ILC2s) via alarmins such as thymic stromal lymphopoietin (TSLP), interleukin (IL)-25, and IL-33. Alarmins play critical roles in the activation of ILC2s and the promotion of T helper (Th)2 cytokine production, leading to the recruitment and activation of eosinophils. In addition to T2-high asthma, non-Th2 mechanisms also exist, leading to neutrophilic asthma phenotypes.⁴,⁵,⁶ Patients with T2-high inflammation can be treated with biologics targeting this inflammation, such as mepolizumab, reslizumab, dupilumab, and omalizumab.⁷ Notably, tezepelumab, which inhibits TSLP, is effective even in patients who are not strictly T2-high.⁸,⁹ TSLP is an upstream cytokine that activates various immune cells, including dendritic cells, T cells, B cells, and ILC2s, contributing to T2-high and non-T2 asthma phenotypes.¹⁰ Understanding these detailed mechanisms and developing targeted therapies could lead to more effective treatment outcomes.¹¹

The distinction between clinical and molecular phenotypes is crucial for a deeper understanding of asthma.¹² Clinical phenotypes are characterized by observable traits, such as symptom severity, frequency of exacerbations, and response to treatment. In contrast, molecular phenotypes are defined by specific molecular and genetic markers that underlie these clinical traits.¹² Identifying molecular phenotypes through RNA sequencing (RNA-seq) allows for a deeper understanding of the disease at the cellular and molecular levels.¹³

In this study, we performed RNA-seq and clustering analysis to identify distinct blood molecular phenotypes of asthma. By comparing these molecular phenotypes with the traditional clinical phenotypes, we aimed to demonstrate the underlying biological mechanisms that contribute to the heterogeneity of asthma. This approach provides insights into the complex pathophysiology of asthma and facilitates more personalized and effective treatment strategies.

MATERIALS AND METHODS

Study design and population

Peripheral blood mononuclear cells (PBMC) RNA-seq data were obtained from 456 adults with asthma in the Cohort for Reality and Evolution of Adult Asthma in Korea (COREA), between January 2005 and December 2020.¹⁴,¹⁵ Of these patients, 23 had severe asthma, while 433 had non-severe asthma. Differentially expressed genes (DEGs) were identified by meta-analysis of publicly available datasets from the Gene Expression Omnibus (GEO) and used to define blood transcriptome-associated clusters (B-TACs). Gene Set Variation Analysis (GSVA) was performed to explore biological pathway activity across clusters. The overall analytic workflow is illustrated in Fig. 1. The study was approved by the Asan Medical Center Institutional Review Board (2019-0376), and all participants provided written informed consent.

Fig. 1. Transcriptomics analysis workflow and validation using COREA and PRISM cohort. The diagram summaries the workflow for the transcriptomics analysis and validation process.

COREA, Cohort for Reality and Evolution of Adult Asthma in Korea; PRISM, Precision Medicine Intervention in Severe Asthma; B-TAC, blood transcriptome-associated cluster; PAM, partition-around-medoids; DEG, differentially expressed gene; GEO, Gene Expression Omnibus.

Public GEO data and DEG meta-analysis

We screened microarray and RNA-seq data from the National Center for Biotechnology Information GEO database¹⁶ in April 2023 using predefined inclusion and exclusion criteria according to Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines (Supplementary Fig. S1).¹⁷ After screening, 8 eligible datasets (6 microarray data, including Unbiased Biomarkers for the Prediction of Respiratory Disease Outcomes (U-BIOPRED), and 2 from RNA-seq) were included for meta-analysis Supplementary Table S1.

DEG and meta-analyses with GEO data

DEG analysis was performed on each microarray or RNA-seq dataset obtained from the GEO using the Limma R package (version 3.42.2),¹⁸ adjusting for age, sex, smoking, and body mass index (BMI) when available. P values from each GEO dataset were combined by random-effects meta-analysis using the inverse-variance method in the ‘meta’ R package (version 4.10.0).¹⁹,²⁰,²¹ A gene was identified as differentially expressed when it had a false discovery rate (FDR)-adjusted P value < 0.05 and appeared in at least 2 out of 8 studies.

RNA-seq

PBMC RNA was extracted from peripheral blood samples of patients with asthma and processed for sequencing following standard protocols. Sequencing reads were quality-controlled, trimmed, and aligned to human genome version 38 (hg38) using HISAT2 (version 2.2.0),¹⁹ and gene-level counts were generated with HTSeq (version 0.12.4) with the UCSC RefGene hg38 GTF file.²⁰ Low-quality or lowly expressed genes were excluded. Data were normalized using the trimmed mean of M values (TMM) methods, and transformed into log-transformed fragments per kilobase of transcript per million mapped reads (FPKM) values for subsequent analysis.

Clustering with RNA-seq data

Clustering was performed using genes selected from meta-DEGs (meta-FDR < 0.05) derived from public GEO data (Fig. 1).²¹ We used the partition-around-medoids (PAM) algorithm with Euclidean distance, which is considered a more robust alternative to the k-means algorithm. To evaluate cluster stability, a consensus clustering method²² was applied using the ConsensusClusterPlus package (version 1.50.0).²³ The optimal number of clusters was determined based on consensus matrix, cumulative distribution function (CDF), and delta area, and cluster quality was evaluated using the Silhouette Index.²⁴

GSVA

GSVA was performed on log-normalized expression data to evaluate differential pathway activity across B-TACs. The analysis was conducted using the GSVA R package (version 1.34.0).²⁵ We focused on 9 gene sets relevant to asthma, as identified by the U-BIOPRED study.²⁶ These gene sets include IL13/Th2, ILC1, ILC2, ILC3, inflammasome, neutrophils, oxidative phosphorylation (OXPHOS), Th17, and aging. Differences in enrichment scores among the 3 B-TACs were evaluated using analysis of variance (ANOVA) followed by t-tests for post-hoc analyses.

Statistical analyses of transcriptome-associated clusters

Differences in clinical variables among the B-TACs groups were analyzed separately in the COREA and PRISM cohorts using appropriate statistical tests according to data distribution and variable type. Continuous variables were compared using ANOVA or nonparametric equivalents, and categorical variables using Pearson χ² or Fisher’s exact test. Analyses were adjusted for relevant covariates (age, sex, BMI, and smoking) when applicable, with multiple comparisons corrected by the FDR method. In the PRISM cohort, drug responsiveness was additionally analyzed in patients who received biologics, with the type of biologics included as a covariate.

Identification of key genes for cluster differentiation

The nearest-shrunken centroid method²⁷ was applied to the log-transformed, TMM-normalized FPKM counts of the 58 genes employed for clustering with the pamr R package (version 1.48.0). The shrinkage parameter was optimized by cross-validation to minimize misclassification error. Genes with non-zero standardized centroids after shrinkage were considered key discriminators among B-TACs.

Validation set

Validation was performed using data from 122 patients with severe asthma from the PRISM cohort. PRISM is a prospective cohort of patients with severe asthma attending asthma clinics in South Korea and the United Kingdom from May 2020 to December 2022.²⁸ Most patients in this study received biologics and were classified as T2-high; moreover, the study exclusively involves Korean participants (Fig. 1, Supplementary Table S2). In the COREA cohort, the median blood eosinophil counts were 320 (140.7–554.5), whereas in the PRISM cohort, the mean levels were higher at 479.7 (215.1–813.9) (Supplementary Table S2). Unlike COREA, PRISM included patients who followed a strict asthma protocol to maximize treatment and ensure adherence, resulting in a cohort composed largely of patients with high blood eosinophilia.

After batch correction of the COREA and PRISM data using the Combat R package,²⁹ B-TAC centroids were recalculated using the 58 genes for clustering. Clinical and molecular characteristics were compared across B-TACs, and pathway activity was assessed by GSVA. Treatment response was primarily assessed on the basis of changes in acute exacerbations and maintenance oral corticosteroids (OCS) dose over 6 months. Responders were defined as patients who achieved at least a 50% reduction in exacerbations or daily OCS dose from baseline. Excellent responders were defined as those with no exacerbations during treatment who also discontinued OCS, whereas super responders were defined as those with no exacerbations during treatment. In addition to these composite responder categories, absolute changes in exacerbation, maintenance OCS dose, asthma control test (ACT) score, and forced expiratory volume in 1 second were also evaluated separately. Detailed protocols, cohort descriptions, and parameter settings for all analyses are provided in Supplementary Data S1.

RESULTS

GEO data meta-analysis and RNA-Seq-based gene clustering

The meta-analysis identified a total of 67 DEGs, 35 upregulated genes, and 32 downregulated genes, with an FDR P value < 0.05 (Fig. 2A, Supplementary Table S3). The top 10 upregulated and downregulated DEGs in the meta-analysis based on the meta-FDR are shown in Table 1. The most significant genes were KCNMA1-AS1 (upregulated) and FANCE (downregulated). Based on the pooled effect sizes, SPANXN3 (upregulated) and LRRN3 (downregulated) revealed the greatest changes in expression. Our results revealed a consistent pattern of gene regulation across the datasets, with a forest plot of the top 5 upregulated and downregulated meta-DEGs depicted in Fig. 2B. This indicated that these genes are fundamental to the molecular mechanisms underlying asthma.

Fig. 2. Meta-analysis across the 8 gene expression omnibus datasets. (A) Visualization of the meta-analysis results with volcano plot. For DEGs (FDR < 0.05), upregulated genes are colored red and downregulated genes are colored blue. The standard error for the combined effect size is depicted as a line. Specifically, the top 5 up- and down-regulated meta-DEGs are presented as gene symbols. The sign consistency indicates the number of datasets used to combine the effect. (B) Forest plots for the top 5 up- and down-regulated meta-DEGs.

DEG, differentially expressed gene; FDR, false discovery rate.

Table 1. Top 10 up- and down-regulated DEGs in meta-analysis based on their meta-FDR.

Gene symbol		Combined effect size	Meta-analysis
			P value	FDR
Top 10 up-regulated DEGs
	KCNMA1-AS1	0.560702179	5.68754E-7	0.00429105612
	OSR2	0.118539063	2.75714E-6	0.0104008341
	SPNS3	0.187685863	5.08496E-6	0.0158811444
	OLIG1	0.272408487	5.61618E-6	0.0158811444
	CLC	0.335047658	6.98999E-6	0.0158811444
	SLC5A9	0.21742922	7.0165E-6	0.0158811444
	CAPN13	0.165480946	9.69694E-6	0.0169791197
	CAPNS2	0.162768578	1.02825E-5	0.0169791197
	IL1RL1	0.16690959	1.04593E-5	0.0169791197
	CLRN3	0.11758814	1.10498E-5	0.0169791197
Top 10 down-regulated DEGs
	FANCE	−0.109616565	2.36664E-7	0.00429105612
	LRRN3	−0.375581329	4.63477E-7	0.00429105612
	NLRP7	−0.196828842	8.0213E-7	0.00453885016
	BMS1P1	−0.123418393	1.42093E-6	0.00643228417
	NDUFA12	−0.123041977	1.12524E-5	0.0169791197
	ANKH	−0.0894907255	1.47197E-5	0.0208227927
	ATP8B1	−0.0942732163	1.87082E-5	0.0211720722
	ARL17B	−0.128445513	2.20737E-5	0.0219411969
	FKBP4	−0.104819739	2.26201E-5	0.0219411969
	LOC100422781	−0.185721254	2.85064E-5	0.0230540435

DEG, differentially expressed gene; FDR, false discovery rate.

Determining optimal clusters in RNA-seq data from COREA cohort

The RNA-seq dataset from the COREA cohort, which underwent a rigorous quality control process, consisted of 456 samples and contained 28,278 genes. To refine our analysis and prevent dimensionality issues, we initially selected the 67 key genes identified in the meta-analysis. However, our dataset contained only 58 genes. Therefore, cluster analysis was performed with these 58 genes. We determined the optimal number of clusters to be 3 using the clustering stability evaluated by consensus clustering. When k = 3, the squares of the consensus matrix were clear (Supplementary Fig. S2A), the mid-section of the consensus CDF was flat (Supplementary Fig. S2B), and the delta value peaked (Supplementary Fig. S2C). This exhibited the highest silhouette index after k = 2 (Supplementary Fig. S2D). Consequently, the data was effectively segmented into 3 B-TACs.

Analysis of key gene expression differences by transcriptome-associated clusters

Using the nearest shrunken centroid method, we identified 11 key genes (ARHGAP36, MMP27, CFTR, CAPN13, OSR2, ENAM, SYN2, POU4F2, GRM6, CLRN3, and IL1RL1) that play a crucial role in differentiating the 3 B-TACs (Supplementary Fig. S3). Notably, these genes were predominantly overexpressed in the B-TAC1. In contrast, B-TAC3 was under-expressed in 4 genes (ARHGAP36, MMP27, CFTR, and CAPN13), with no significant trends observed for the remaining genes. Interestingly, the B-TAC2 group exhibited a standardized centroid value of 0 for all 11 genes, suggesting a neutral expression profile and positioning B-TAC2 as an intermediate phenotype between B-TAC1 and B-TAC3 (Fig. 3A). The clinical phenotypic trends in absolute blood eosinophil counts among the 3 B-TAC groups are depicted in Fig. 3B.

Fig. 3. Gene expression and T2 clinical phenotype differences among the 3 transcriptome-associated clusters (B-TAC1, B-TAC2, and B-TAC3). (A) Heatmap of expression levels for 58 significant genes, identified by meta-analysis and present in our dataset, from 456 individuals. The genes were categorized as upregulated (light pink) or downregulated (light blue) in asthma, and they were sorted based on their false discovery rate within each category. (B) Boxplots of absolute blood eosinophil counts (log-transformed for visualization).

B-TAC, blood transcriptome-associated cluster.

Clinical phenotypic difference between 3 transcriptome-associated clusters

The clinical characteristics of the 3 B-TAC groups are presented in Table 2. The median age was 53 years for patients of B-TAC1, 48 for B-TAC2, and 55 for B-TAC3, with B-TAC2 having the lowest median age. The age at asthma diagnosis was the lowest in the B-TAC2 group at 44 years, compared to 50 years for B-TAC1 and 46 years for B-TAC3 groups, although these differences were not statistically significant (P = 0.093). The percentage of patients considered by physicians to have asthma-chronic obstructive pulmonary disease (COPD) overlap was highest in the B-TAC2 group (9.2%) and lowest in the B-TAC1 group (5.6%), although these differences were not statistically significant. There were significant differences in BMI among the B-TAC groups, with patients in B-TAC2 group having a relatively low BMI of 23.70 kg/m² (21.90–26.70 kg/m²). There were no significant differences in sex or smoking history between the groups. The percentage of patients who reported having comorbidities such as hypertension and diabetes was marginally higher in the B-TAC3 group; however, the difference was not statistically significant. Additionally, there were no differences in the prevalence of conditions such as allergic rhinitis, chronic sinusitis, or atopic dermatitis between the groups. Although the difference was not statistically significant, the B-TAC3 group revealed a lower tendency to report atopy. Lung function was the highest in B-TAC1 group, while B-TAC2 and B-TAC3 groups revealed a lower function; however, the difference was not statistically significant. Absolute blood eosinophil levels (cells/μL) were significantly elevated in both B-TAC1 and B-TAC2 groups, with respective values of 336.40 cells/μL (164.75–543.30 cells/μL) and 369.70 cells/μL (153.30–642.40 cells/μL) (P = 0.012). Total IgE and induced sputum eosinophil percentages were typically higher in the B-TAC1 and B-TAC2 groups than in the B-TAC3 group; however, the differences were not statistically significant. Fractional exhaled nitric oxide (FeNO) levels were higher in the B-TAC1 group (52 ppb), whereas the B-TAC2 and B-TAC3 groups revealed marginally lower levels at 37 and 40 ppb, respectively, although these differences were not statistically significant. There were no differences between the groups in the use of inhalers or leukotriene receptor antagonist medications; however, the proportion of patients on maintenance oral steroids was significantly higher in the B-TAC3 group (32.69%, P = 0.031).

Table 2. Clinical characteristics of the 3 transcriptome-associated clusters (COREA).

Variable		B-TAC1 (n = 125)	B-TAC2 (n = 196)	B-TAC3 (n = 135)	P value
Age (yr)		53.00 (40.00–63.00)	48.00 (35.50–59.50)	55.00 (37.50–62.00)	0.072
Age of asthma diagnosis (yr)		50.00 (36.00–60.00)	44.00 (32.00–55.00)	46.00 (32.00–59.00)	0.093
Asthma-COPD overlap*		7 (5.6%)	18 (9.2%)	10 (7.4%)	0.496
Sex					0.064
	Female	79 (63.20%)	123 (62.76%)	69 (51.11%)
	Male	46 (36.80%)	73 (37.24%)	66 (48.89%)
BMI (kg/m2)		24.30 (22.10–27.30)	23.70 (21.90–26.70)	25.80 (23.00–27.95)	0.001
Smoking					0.699
	Current	11 (8.80%)	28 (14.29%)	16 (11.85%)
	Ex-smoker	39 (31.20%)	59 (30.10%)	41 (30.37%)
	Non-smoker	75 (60.00%)	109 (55.61%)	78 (57.78%)
Hypertension					0.150
	No	122 (97.60%)	193 (98.47%)	128 (94.81%)
	Yes	3 (2.40%)	3 (1.53%)	7 (5.19%)
Diabetes mellitus					0.139
	No	123 (98.40%)	194 (98.98%)	129 (95.56%)
	Yes	2 (1.60%)	2 (1.02%)	6 (4.44%)
Allergic rhinitis					0.356
	No	78 (62.40%)	107 (54.59%)	75 (55.56%)
	Yes	47 (37.60%)	89 (45.41%)	60 (44.44%)
Chronic rhinosinusitis					0.307
	No	124 (99.20%)	192 (97.96%)	130 (96.30%)
	Yes	1 (0.80%)	4 (2.04%)	5 (3.70%)
Atopic dermatitis					0.469
	No	118 (94.40%)	181 (92.35%)	129 (95.56%)
	Yes	7 (5.60%)	15 (7.65%)	6 (4.44%)
Atopy					0.261
	No	18 (25.00%)	28 (22.58%)	28 (32.56%)
	Yes	54 (75.00%)	96 (77.42%)	58 (67.44%)
Lung function†
	FEV1 predicted value (%)	71.00 (57.00–80.41)	66.00 (54.50–77.00)	66.00 (56.50–77.92)	0.666
	FVC predicted value (%)	80.00 (70.00–90.22)	76.00 (68.00–89.00)	74.00 (67.00–80.50)	0.310
	FEV1/FVC ratio	0.69 (0.60–0.74)	0.68 (0.61–0.79)	0.71 (0.61–0.79)	0.979
Blood IgE total level (kIU/L)		215.00 (88.00–432.00)	226.00 (78.90–531.00)	180.00 (60.60–516.50)	0.440
Absolute blood eosinophils (/μL)		336.40 (164.75–543.30)	369.70 (153.30–642.40)	279.65 (126.50–445.20)	0.012
Induced sputum eosinophil (%)		3.00 (0.00–13.50)	3.00 (0.00–18.00)	1.50 (0.00–8.00)	0.159
FeNO (ppb)		52.00 (24.00–95.00)	37.00 (22.50–80.50)	40.00 (23.00–71.00)	0.334
ICS dose					0.142
	Low dose	62 (51.67%)	86 (45.50%)	56 (42.42%)
	Medium dose	37 (30.83%)	48 (25.40%)	36 (27.27%)
	High dose	21 (17.50%)	55 (29.10%)	40 (30.30%)
Medication
	ICS/LABA	118 (94.40%)	187 (95.41%)	124 (91.85%)	0.397
	ICS/LABA/LAMA	22 (17.60%)	26 (13.27%)	20 (14.81%)	0.568
LTRA					0.512
	No	35 (59.32%)	40 (50.00%)	29 (50.88%)
	Yes	24 (40.68%)	40 (50.00%)	28 (49.12%)
Oral corticosteroids maintenance					0.031
	No	46 (80.70%)	64 (86.49%)	35 (67.31%)
	Yes	11 (19.30%)	10 (13.51%)	17 (32.69%)
Asthma control status					0.612
	Controlled	25 (20.00%)	46 (23.47%)	33 (24.44%)
	Partly controlled	69 (55.20%)	115 (58.67%)	74 (54.81%)
	Uncontrolled	31 (24.80%)	35 (17.86%)	28 (20.74%)
History of asthma exacerbation (ever)					0.400
	Yes (≥ 1 exacerbation)	34 (27.20%)	65 (33.16%)	37 (27.41%)
	No	91 (72.80%)	131 (66.84%)	98 (72.59%)
History of asthma exacerbation (previous 1 year)					0.828
	Yes (≥ 1 exacerbation)	25 (20.00%)	40 (20.41%)	24 (17.78%)
	No	100 (80.00%)	156 (79.59%)	111 (82.22%)

Values are presented as median (1st–3rd quartiles) or number (%).

COREA, Cohort for Reality and Evolution of Adult Asthma in Korea; B-TAC, blood transcriptome-associated cluster; COPD, chronic obstructive pulmonary disease; BMI, body mass index; FEV1, forced expiratory volume in 1 second; FVC, forced vital capacity; IgE, immunoglobulin E; FeNO, fractional exhaled nitric oxide; ICS, inhaled corticosteroids; LABA, long-acting beta-agonist; LAMA, long-acting muscarinic antagonist; LTRA, leukotriene receptor antagonist.

^*Physician diagnosis asthma-COPD overlap; ^†Post-bronchodilator.

GSVA

To determine the gene sets that were active in each B-TAC group, we conducted a GSVA on 9 gene sets related to asthma (Fig. 4). B-TAC1 exhibited the highest expression of IL13_Th2 and ILC2 gene sets, which are active in the Th2-immune response. In contrast, B-TAC3 exhibited the highest expression of Th17, ILC1, and neutrophil gene sets, which are active in non-Th2 immune responses (Th1/Th17). Although there was a significant difference in ILC3 expression among the 3 clusters, the difference in the mean values was < 0.1, and thus was not included in the interpretation. The B-TAC2 group displayed an intermediate expression pattern for all the above gene sets, suggesting that B-TAC2 is characterised by mixed immune responses comprising Th2 and Th1/Th17 immune responses.

Fig. 4. GSVA for 9 different gene signatures associated with asthma. Boxplots depicting the difference in enrichment scores for the 9 gene sets associated with asthma among the 3 transcriptome-associated clusters (B-TAC1, B-TAC2, and B-TAC3).

GSVA, Gene Set Variation Analysis; B-TAC, blood transcriptome-associated cluster; IL, interleukin; Th, T-helper cell; ILC, innate lymphoid cell; OXPHOS, oxidative phosphorylation.

^*P < 0.05; ^**P < 0.01; ^***P < 0.001; ^****P < 0.00001.

Validation from the PRISM cohort

Using the RNA-seq clustering of the 122 patients included in the PRISM validation cohort, the patients were categorized into B-TAC1 (n = 20) and B-TAC2 (n = 102) groups, with no patients in the B-TAC3 group (Supplementary Fig. S4). Most patients in B-TAC1 received biologic therapy (5% received conventional treatment), whereas 26.5% of patients in the B-TAC2 group received conventional treatment (Fig. 5A). The distribution of clinical phenotypes, as perceived by clinicians, showed that in the B-TAC1 group, only one patient (5%) was considered to have a T2-low phenotype, whereas in the B-TAC2 group, 12 (12%) patients were considered to have a T2-low phenotype (Fig. 5B). The baseline clinical characteristics of the groups are presented in Supplementary Table S4. Both B-TAC1 and B-TAC2 groups had blood eosinophil counts > 300 cells/μL, and no statistically significant differences were observed between the groups (Fig. 5C). When comparing the treatment effects between the B-TAC1 and B-TAC2 groups, no statistically significant differences were observed in the improvement of lung function, change in exacerbation, change in maintenance OCS dose, or distribution of responders. However, there was a significant difference in ACT scores between the groups. The B-TAC1 group showed a significant improvement in the ACT score, with a 7-point increase, compared to the 4-point improvement in the B-TAC2 group (Table 3, Fig. 5D). GSVA was also performed, and both B-TAC1 and B-TAC2 revealed patterns similar to those observed in the previous COREA cohort analysis (Fig. 4).

Fig. 5. The differences in clinical phenotypes and clinical characteristics according to molecular phenotypes in the validation set (PRISM cohort). (A) The distribution of treatment medications according to the B-TAC groups reveals that a higher proportion of biologics use is observed in the B-TAC1, whereas a relatively higher proportion of patients receiving conventional treatment is observed in the B-TAC2. (B) The distribution of clinical phenotypes as perceived by clinicians reveals that in the B-TAC1, only 1 patient (5%) was considered to have a T2-low phenotype, while in the B-TAC2, 12% of patients were thought to have a T2-low phenotype. The inner circle represents the B-TAC1, whereas the outer circle represents the B-TAC2. (C) Comparison of absolute blood eosinophil counts between B-TAC groups. (D) The improvement in ACT scores between B-TAC groups after six months of treatment.

PRISM, Precision Medicine Intervention in Severe Asthma; B-TAC, blood transcriptome-associated cluster; ACT, asthma control test.

Table 3. Treatment responsiveness between molecular phenotypes in the validation set (PRISM cohort).

Variable	B-TAC1 (n = 19)	B-TAC2 (n = 75)	P value
Responders	17 (89.5%)	70 (93.3%)	0.709
Excellent responders	13 (68.4%)	55 (73.3%)	0.647
Super responders	13 (68.4%)	52 (69.3%)	0.893
Increase in ACT score	7 (3.5 to 9.5)	4 (2.00 to 8.00)	0.048
Increase in FEV1 (mL)	190.0 (0.0 to 430.0)	265.0 (62.5 to 512.5)	0.562
Increase in FEV1 (%)	7.0 (0.5 to 15.0)	9.5 (2.0 to 16.0)	0.783
Change in exacerbation	−2 (−5 to 0)	−2 (−5 to 0)	0.533
Change in OCS dose (mg/day)	0 (0 to 0)	0 (−2.5 to 0)	0.999

Values are presented as median (1st–3rd quartiles) or number (%).

PRISM, Precision Medicine Intervention in Severe Asthma; B-TAC, blood transcriptome-associated cluster; ACT, asthma control test; FEV1, forced expiratory volume in 1 second; OCS, oral corticosteroids.

DISCUSSION

In this study, we explored the molecular phenotypes of asthma by analyzing RNA-sequencing data from PBMCs, which revealed 3 B-TACs within the COREA cohort. B-TAC1 was characterized by activation of the IL-13/Th2 and ILC2 gene sets, with the clinical phenotype displaying high blood eosinophil levels and elevated FeNO, indicative of T2-high features. Additionally, the B-TAC1 group exhibited increased expression of genes associated with enzymes expressed in eosinophils and basophils, such as CLC, and genes related to IL-33 receptors (IL1RL1/ST2).³⁰,³¹ Conversely, B-TAC3 revealed a lack of activity in the IL-13/Th2 and ILC2 gene sets. Instead, B-TAC3 was marked by the activation of gene sets associated with neutrophils, inflammasomes, ILC1, OXPHOS, aging, and Th1, correlating with lower levels of blood eosinophils and reduced FeNO, indicative of low T2 features. Consequently, within the PRISM cohort, which primarily consisted of patients exhibiting the clinical T2-high phenotype and who were candidates for biological treatment, there were no individuals corresponding to B-TAC3.

Importantly, our main finding is that even among patients who appear clinically similar based on conventional T2-related markers, particularly elevated blood eosinophil counts, distinct molecular phenotypes can be identified by blood transcriptomic clustering. In particular, B-TAC1 and B-TAC2 both showed eosinophilic/T2-high clinical features, yet differed in their transcriptomic signatures and in their tendency toward treatment response, indicating that eosinophilia alone does not fully reflect the underlying inflammatory endotypes. Notably, the B-TAC2 group was identified as a mixed group, exhibiting characteristics of both the B-TAC1 and B-TAC3 groups. B-TAC2, despite revealing high blood eosinophil counts, which clinically suggests a T2-high phenotype, was identified with a molecular phenotype that differs from that of B-TAC1. Notably, B-TAC2 displayed the most significant downregulation of ILC3 gene set. These characteristics were also observed in the validation cohort.

These results indicated that even in patients who clinically exhibit increased blood eosinophil counts, the underlying molecular mechanisms may differ, supporting the diversity of asthma. While both the B-TAC1 and B-TAC2 groups showed clinically elevated blood eosinophils, physicians have observed that COPD overlap is more prevalent in the B-TAC2 group, and the proportion of patients considered T2-high was marginally lower in B-TAC2 than in B-TAC1. Additionally, the B-TAC2 group had a higher propensity to opt for conventional treatments than the biologics group. After 6 months, the ACT score did not improve significantly in the B-TAC2 group compared to that in the B-TAC1 group. There were non-responders in the treatment of T2 inflammation with biologics, and these patients typically exhibit mixed-type inflammation, persistent airflow obstruction, and have numerous comorbid conditions.³²,³³,³⁴ In clinical practice, treatment individualization in severe asthma still relies mainly on established biomarkers and clinical features. Our findings suggest that blood transcriptomic categorization may complement these approaches by identifying molecular heterogeneity among patients with apparently similar T2-high features. Our results may help explain the molecular mechanisms of these patients.

The U-BIOPRED group previously conducted similar analyses using sputum samples, in which TAC1 and TAC2 were primarily categorized as Th2-high and non-Th2, respectively.²⁶ The differences in results are considered to be primarily owing to the type of sample used. Compared to localized sputum samples, the blood samples used in our study may reflect more complex systemic inflammation. This difference in sample type may also influence the treatment-response signals that can be detected. Blood-based transcriptomic profiles may better reflect systemic inflammatory states and circulating immune activation, whereas sputum or airway samples may more directly capture local airway inflammation relevant to airway-specific treatment responses. Therefore, future studies evaluating biologic responsiveness may benefit from integrating both systemic and airway-local biomarkers to improve endotyping and biomarker selection. This could be a potential limitation of the previous study. However, in our study, the use of blood samples as opposed to sputum samples used in traditional analyses provided a broader systemic perspective, potentially leading to more comprehensive biomarker identification. This approach supports the development of targeted therapies that could improve outcomes in patients with asthma who do not respond to conventional treatments.

This study has the following limitations. First, the findings are based on data from the COREA and PRISM cohorts, which may limit generalizability to other populations with varying environmental exposures. In addition, the PRISM validation cohort was enriched for Korean patients with T2-high severe asthma, with high blood eosinophil counts and no representation of B-TAC3. Furthermore, the B-TAC groups were imbalanced in size, with a substantially smaller number of patients in B-TAC1 than in B-TAC2. Therefore, findings related to biologic responsiveness should be interpreted cautiously, and their generalizability to broader severe asthma populations may be limited. Expanding future research to include diverse, multinational cohorts could help address this limitation. Second, while peripheral blood samples were used to capture systemic inflammation, they may not fully represent localized inflammation in the respiratory tract. Future studies incorporating sputum or respiratory tissue samples could provide a more comprehensive understanding of asthma pathology. Third, further validation of the biomarkers identified in this study is necessary to confirm their predictive value for clinical outcomes. Longitudinal studies evaluating their effectiveness in real-world clinical settings would enhance their translational potential. Finally, biologic-specific subgroup analyses were limited by the small sample size of patients in each treatment category, particularly within B-TAC1. Therefore, potential differences in treatment response according to biologic mechanism could not be robustly evaluated and should be addressed in larger future studies. Nevertheless, these limitations do not diminish the impact of our findings.

In conclusion, our detailed examination of B-TACs in asthma differentiated by distinct molecular and clinical phenotypes underscores the critical role of precision medicine in the management of complex diseases. Our findings revealed that even among patients with similar clinical profiles, underlying molecular differences can significantly impact treatment responses. Notably, the identification of mixed inflammatory profiles and the delineation between systemic and localized inflammation highlights the need for individualized treatment approaches.

SUPPLEMENTARY MATERIALS

Supplementary Data S1

Study methods

Supplementary Table S1

The 8 GEO datasets used in the meta-analysis

Supplementary Table S2

Characteristics of the training and validation sets

Supplementary Table S3

Sixty-seven DEGs identified in the meta-analysis of 8 datasets

Supplementary Table S4

Clinical characteristics of the 2 B-TACs in PRISM

Supplementary Fig. S1

Dataset selection process for meta-analysis. The data selection process from the GEO database. GEO, Gene Expression Omnibus; RNA-seq, RNA-sequencing; lncRNA, long non-coding RNA; PBMC, peripheral blood mononuclear cells.

Supplementary Fig. S2

Consensus clustering using the PAM algorithm. (A) Consensus matrix for 3 clusters. (B) Consensus CDF of k = 2 to k = 10. (C) Internal index of cluster quality. (D) Delta area under the CDF curve. Given these plots, we decided to select k = 3. PAM, partition-around-medoids; CDF, cumulative distribution function.

Supplementary Fig. S3

The nearest shrunken centroid method. (A) The classifier for the three clusters was trained using cross-validation. The shrinkage parameter threshold of 11.013 (indicated by a black dashed line) was selected because it significantly reduced the total misclassification error, narrowing down the classifier to 11 genes. (B) Centroid profile of the 11 genes identified by cross-validation. It displays the standardized centroid values of the 11 genes identified through cross-validation across each cluster. Positive values (blue) indicate a gene’s up-regulation in a cluster, while negative values (red) signify down-regulation. The magnitude of these values reflects the gene’s importance in distinguishing the cluster, with higher absolute values highlighting key biological differences. B-TAC, blood transcriptome-associated cluster.

Supplementary Fig. S4

Multi-dimensional scaling plot showing the distribution of patients from the COREA and PRISM cohorts based on B-TAC classifications. Each point represents an individual patient, color-coded by B-TAC type (B-TAC1, B-TAC2, B-TAC3) and cohort. The ellipses represent the 95% confidence intervals for each B-TAC type within each cohort, illustrating the similarity or separation between patient groups in multidimensional space.