Plasma Metabolomes Identify Distinct Asthma Metabotypes: Findings From the COREA Cohort

ABSTRACT

Background

Phenotype‐based asthma classification has limitations that have motivated endotype‐based approaches grounded in pathophysiology. However, existing biomarkers offer limited specificity and clinical accessibility. Plasma metabolomics, which reflects disease‐specific metabolic states, offers a promising strategy for asthma classification. This study aimed to identify metabolic subgroups (metabotypes) of adult asthma using plasma metabolomics, with potential implications for personalized treatment.

Methods

Plasma samples from 407 patients with asthma in the Cohort for Reality and Evolution of Adult Asthma in Korea (COREA) were analyzed using Biocrates AbsoluteIDQ p400 HR kit with liquid chromatography‐mass spectrometry. After preprocessing, 281 metabolites were natural‐log‐transformed, standardized, and partitioned by k‐means clustering. The number of clusters was determined by a multi‐criteria assessment combining cluster validity indices, consensus clustering, stability analysis, and cross‐algorithm agreement.

Results

Four metabotypes with distinct lipid‐class signatures were identified. Group 1 (n = 123) was characterized by elevated ether‐linked phosphatidylcholines, comprising younger patients with the earliest symptom onset and female predominance. Group 2 (n = 53) showed elevated lyso‐phosphatidylcholines and altered amino acid metabolism (elevated glutamate, reduced glutamine), representing a metabolically intermediate, non‐T2‐high subgroup. Group 3 (n = 158) exhibited globally reduced sphingomyelins and broadly lower phosphatidylcholines in middle‐aged, non‐obese patients. Group 4 (n = 73) showed markedly elevated triacylglycerols (TG) and diacylglycerols (DG) with the highest body mass index (BMI), consistent with a non‐T2, obesity‐related metabotype. The TG/DG signature of Group 4 remained robust after adjustment for BMI, age, and sex.

Conclusion

Plasma metabolomics identifies clinically meaningful asthma metabotypes, supporting integration of metabolomic profiling into personalized asthma management.

This study aimed to identify metabolic subgroups (metabotypes) of adult asthma using plasma metabolomics, with potential implications for personalized treatment. Four metabotypes with distinct lipid‐class signatures were identified. Group 1 had elevated ether‐linked phosphatidylcholines with young, female‐predominant early onset; Group 2 showed elevated lyso‐phosphatidylcholines and altered amino acid metabolism; Group 3 exhibited reduced sphingomyelins and phosphatidylcholines in middle‐aged, non‐obese patients; and Group 4 demonstrated elevated triacylglycerols and diacylglycerols with the highest BMI.

1. Introduction

Asthma is a heterogeneous disease characterized by multiple phenotypes and endotypes [1]. Recent studies have focused on implementing endotype‐based classification to provide personalized treatment tailored to patients with diverse asthma profiles [2, 3]. Understanding the mechanistic underpinnings of inflammatory endotypes, particularly type 2 (T2) inflammation, is critical for guiding treatment. However, these mechanisms remain incompletely understood, and the associated biomarkers, including the absolute eosinophil count (AEC) in blood, the percentage of eosinophils in sputum, and fractional exhaled nitric oxide (FeNO), exhibit low specificity and high variability [4, 5].

Because metabolite levels are sensitive to physiological state, they effectively reflect phenotypes, making metabolomics a widely used tool for biomarker development [6]. Their sensitivity to disease mechanisms also allows metabolomics to be utilized for endotype classification in asthma [7]. Moreover, metabolomics requires less invasive sampling and supports high‐throughput analysis with minimal sample volumes, rendering metabolomics‐based metabotyping—defined here as the classification of patients into subgroups reflecting distinct plasma metabolic profiles—particularly well‐suited to clinical settings.

Several studies have explored asthma metabotypes using metabolomics. Comhair et al. reported significant differences in plasma metabolites between asthma and control groups [7]. Reinke et al. noted that asthma severity and inhaled corticosteroid (ICS) use corresponded with alterations in metabolic markers, including steroid hormones [8]. Most studies have focused on case–control comparisons, whereas metabolomics‐based clustering within asthma populations remains limited, with even fewer studies conducted in East Asian populations.

This study aims to identify novel metabotypes in patients with asthma in the cohort for reality and evolution of adult asthma in Korea (COREA) through unsupervised clustering. These findings may, in turn, inform personalized treatment strategies tailored to the metabolic and clinical features of each cluster.

2. Materials and Methods

The overall study workflow is summarized in Figure S1.

2.1. Study Population

A total of 407 adult patients with asthma from the COREA cohort with available plasma samples and medical records were enrolled [9]. Data were collected at 27 hospitals nationwide between January 2005 and December 2020. The inclusion criteria were (1) age > 18 years; (2) symptoms such as dyspnea, wheezing, or cough for more than 3 months without acute exacerbation at the time of enrollment; and (3) airway hyperresponsiveness, as indicated by either a 20% reduction in forced expiratory volume in 1 s (FEV1) with a methacholine dose of 16 mg/mL (PC20) or airway reversibility, as indicated by an increase in FEV1 of > 12% (and at least 200 mL) following the inhalation of a short‐acting β‐agonist. All patients were diagnosed with asthma by allergists or pulmonologists. Patients were excluded if active or acute pulmonary infection, severe lung damage, bronchiectasis, or a history of lung resection was identified. Blood samples were collected at each participating center, and all samples were processed according to a standardized protocol. Induced sputum samples were obtained by the inhalation of nebulized sterile saline solution followed by coughing and expectoration of airway secretions. This study was approved by the institutional review board (IRB) of Asan Medical Center (IRB No. 2019‐0376) and conducted in accordance with the declaration of Helsinki. All patients provided written informed consent.

2.2. Plasma Metabolite Profiling

Frozen plasma samples were thawed on ice and prepared with the AbsoluteIDQ p400 HR kit (Biocrates Life Sciences AG, Innsbruck, Austria) according to the manufacturer's instructions. To monitor and correct for batch‐to‐batch variability, pooled quality control (QC) samples were prepared by combining equal volumes from a subset of analytical samples and processed under the same protocol. A total of 408 metabolites were measured using an Ultimate 3000 ultra‐performance liquid chromatograph coupled with a Q Exactive Plus hybrid quadrupole‐orbitrap mass spectrometer (Thermo Fisher Scientific, MA, USA). Of 408 metabolites measured, 42 were quantified in liquid chromatography (LC) mode, and 366 were semi‐quantified in flow injection analysis (FIA) mode.

Raw data acquired in LC mode were processed using the Xcalibur Software (Thermo Fisher Scientific). LC‐processed data and raw FIA data were simultaneously processed with MetIDQ Oxygen v.2893 (Biocrates Life Sciences AG) and exported as micromolar values. Metabolites detected below the limit of detection (LOD) in more than 12 out of 18 pooled QC samples, or in more than 70% of analytical samples, were excluded; 281 metabolites were retained. Values below LOD were imputed as one‐fifth of the minimum positive value per metabolite, an approach consistent with established metabolomics practice that avoids zero‐imputation artifacts. FIA‐mode concentrations were normalized using the QC‐based random forest signal correction (QC‐RFSC) method [10]; LC‐mode concentrations were retained as absolute values without further normalization.

2.3. Cluster Analysis

Prior to clustering, the 281‐metabolite concentration matrix was natural‐log‐transformed using ln(1 + x) and z‐score standardized per metabolite. Patients were grouped using the k‐means algorithm (scikit‐learn) with k‐means++ initialization, 50 random restarts (n_init = 50), a maximum of 500 iterations, and a fixed random seed; the partition with the lowest within‐cluster inertia was retained. The optimal number of clusters (k = 4) was determined by integrating principal component analysis (PCA), quantitative cluster validity indices (silhouette, Calinski–Harabasz, Davies–Bouldin, and the gap statistic), bootstrap consensus clustering, partition stability under random‐seed perturbation, agreement with four alternative clustering algorithms (Gaussian mixture model, k‐medoids, spectral clustering, and agglomerative clustering with Ward linkage), and clinical interpretability; full details are provided in Supplementary Methods, “Assessment of cluster validity”.

2.4. Statistical Analysis

For the 281‐metabolite panel, an omnibus test (one‐way ANOVA or Kruskal–Wallis, selected by within‐cluster Shapiro–Wilk normality) was performed per metabolite, with eta‐squared (η ²) or epsilon‐squared (ε²) as the effect size. Cluster‐discriminative metabolites were further identified per cluster by one‐vs‐rest Welch's t‐test with Cohen's d as a standardized effect size, defined as those with |Cohen's d| ≥ 0.8 and Benjamini–Hochberg false discovery rate (BH‐FDR)‐adjusted p < 0.05. Class‐level enrichment within each cluster's discriminative set was assessed by one‐sided Fisher's exact tests against the full 281‐metabolite panel. Clinical variables across the four metabotypes were compared using one‐way ANOVA with Tukey's HSD, Kruskal–Wallis with Mann–Whitney U and Bonferroni correction, or chi‐square/Fisher's exact tests, with η ², ε², or Cramer's V as effect sizes. Clinical variables with missing data were analyzed under complete‐case analysis. All p‐values were adjusted using the BH‐FDR method within each analysis panel. To address potential confounding, three complementary multivariable analyses were performed: (i) multinomial logistic regression of cluster membership on age, sex, BMI, log‐eosinophil count, smoking status, and ICS use (Group 3 as reference); (ii) per‐metabolite ordinary least‐squares regression testing whether discriminative signatures persisted after adjustment for the same covariates; and (iii) a Group 4‐specific lipid analysis testing robustness to BMI, age, and sex adjustment. Full details are provided in Supplementary Methods, “Confounding adjustment”. No outlier filtering was applied; log transformation and non‐parametric testing reduced the impact of extreme values. All analyses were performed using Python 3.14.4 with scikit‐learn 1.8.0, statsmodels 0.14.6, and SciPy 1.17.

2.5. Declaration of Generative AI and AI‐Assisted Technologies in the Manuscript Preparation Process

During the preparation of this work, the authors used ChatGPT‐5.2 and Claude Sonnet 4.6 in order to assist with English editing and language refinement, and Claude Code Opus 4.7 for code review and revision. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.

3. Results

3.1. Distribution of Plasma Metabolites by Metabotype Clustering

After batch‐effect correction of plasma metabolite levels (Figure S2), the 407 patients with asthma were grouped into four clusters on the basis of 281 plasma metabolite levels using k‐means clustering. The optimal algorithm and the number of clusters (k = 4) were determined by integrating PCA, quantitative cluster validity indices, consensus clustering, hyperparameter stability analysis, agreement across alternative clustering algorithms, and clinical interpretability (Figures S3–S8; Supplementary Methods). The numbers of patients in each cluster were 123 (Group 1), 53 (Group 2), 158 (Group 3), and 73 (Group 4). PCA showed a distinct distribution between the cluster groups (Figure S8A–B), and per‐metabolite omnibus testing revealed that 92.9% (261/281) of metabolites differed significantly across clusters at FDR‐adjusted p < 0.05 (Figure S8C).

To characterize the metabolite distribution underlying each metabotype, we identified cluster‐discriminative metabolites, on the basis of standardized effect size (|Cohen's d| ≥ 0.8) and statistical significance (FDR‐adjusted p < 0.05) when comparing each cluster against the remaining three. Each metabotype was characterized by a distinct lipid‐class‐dominant signature (Figure 1 and Figure S9): Group 1 by elevated ether‐linked phosphatidylcholines (PC‐O), Group 2 by elevated lyso‐phosphatidylcholines (LPC) and ether‐linked LPC (LPC‐O) together with several altered amino acids, Group 3 by globally reduced sphingomyelins (SM) and broadly lower phosphatidylcholines (PC), and Group 4 by markedly elevated triacylglycerols (TG) and diacylglycerols (DG). Fisher's exact enrichment analysis confirmed selective over‐representation of the dominant lipid classes (PC‐O in Group 1, LPC/LPC‐O in Group 2, SM in Group 3, and TG/DG in Group 4; all FDR‐adjusted p < 0.05; Table S1, Figure S10). To rule out confounding by clinical covariates known to influence circulating metabolites, we re‐tested every cluster‐metabolite contrast after adjustment for age, sex, body mass index (BMI), smoking status, and ICS use. The majority of unadjusted‐significant metabolites retained FDR‐adjusted p < 0.05 (Group 1, 83%; Group 2, 93%; Group 3, 97%; Group 4, 97%, Table S2). For Group 4 specifically, 161 of 166 lipids significant in the unadjusted comparison remained significant after additional adjustment for BMI, age, and sex, with median attenuation of 8.0% (Table S3, Figure S11), indicating that the TG‐loaded lipid signature of Group 4 is not a passive consequence of higher adiposity.

FIGURE 1.

Standardized mean plasma metabolite levels across the four cluster groups identified by k‐means clustering. Radar plots display the mean standardized levels (z‐scores) of plasma metabolites in each group. Each dot represents the mean z‐score of a metabolite in a given cluster, and lines connect successive metabolite means within each cluster for visualization purposes. Metabolites are grouped by class: AA, amino acids; AC, acylcarnitines; BGA, biogenic amines; Cer, ceramides; CE, cholesteryl esters; DG, diacylglycerols; DG‐O, ether‐linked diacylglycerols; Hex, hexose; LPC, lysophosphatidylcholines; LPC‐O, ether‐linked lysophosphatidylcholines; PC, phosphatidylcholines; PC‐O, ether‐linked phosphatidylcholines; SM, sphingomyelins; TG, triacylglycerols.

3.2. Metabolic Characteristics of Metabotype Clusters

The four metabotypes were defined by distinct lipid‐class signatures (Figure 1). Group 1 showed a marked elevation in PC‐O species: 19 of its 23 cluster‐discriminative metabolites belonged to the PC‐O class (Fisher's exact OR 89.5; FDR‐adjusted p < 0.001; Table S1), with PC‐O(40:7), PC‐O(40:6), and PC‐O(40:5) showing the largest discrimination (|Cohen's d| > 1.1; Table S4). Three SM species, most prominently SM(34:1), were also elevated. Group 2 was dominated by LPC and LPC‐O species (11 LPC and 4 LPC‐O among its 35 discriminative metabolites; Fisher's exact OR 37.1 and 31.6; FDR‐adjusted p < 0.001 and p = 0.005), with the largest discrimination among lipid species (|Cohen's d| > 1.5). The signature also extended into amino acid metabolism: eight amino acids reached the discriminative threshold, with glutamate showing the strongest elevation (|Cohen's d| = 2.409) and glutamine the strongest reduction (|Cohen's d| = 1.774).

In contrast to the elevation‐driven signatures of Groups 1 and 2, Group 3 was defined by a broadly suppressed lipid profile, with all 41 cluster‐discriminative metabolites reduced relative to the remaining clusters. SM drove this signature (13 of 30 panel SM species; Fisher's exact OR 6.1, FDR‐adjusted p < 0.001), with SM(41:2), SM(41:1), and SM(38:1) showing the largest discrimination (|Cohen's d| > 1.17). Reductions also extended to 15 PC and 7 PC‐O species, although these broader classes did not reach class‐level enrichment significance (FDR‐adjusted p = 0.140 and 0.600). Group 4 exhibited the most extensive metabolic perturbation, with 80 cluster‐discriminative metabolites and an overwhelmingly upward direction (77 of 80 elevated). The signature was dominated by neutral lipids: TG and DG were both significantly enriched (Fisher's exact OR 16.9 and 4.1; FDR‐adjusted p < 0.001 and p = 0.048), with TG(50:3), TG(51:3), and TG(50:2) showing exceptionally large discrimination (|Cohen's d| > 2.0). The Group 4 signature was further characterized by a shift toward shorter total acyl carbon numbers for TG (median 51 vs. panel median 52), PC (35.5 vs. 37), and DG (36 vs. 38), and toward longer chains for SM (41 vs. 38; all FDR‐adjusted p < 0.05) (Figure S12).

3.3. Clinical Characteristics of Metabotype Clusters

The four metabotypes also differed substantially in demographic and clinical features (Table 1). Group 1 included the youngest participants (mean age 44.9 years) and the earliest symptom onset (mean 38.9 years), whereas Groups 2–4 were comparable in age (mean 53–54 years, FDR‐adjusted p < 0.001). Sex distribution differed across groups, with females predominating in Group 1 (75%) compared with a more balanced distribution in Group 3 (50%; FDR‐adjusted p = 0.006). BMI showed a graded increase from Group 1 (23.8 kg/m²) to Group 4 (26.1 kg/m²; FDR‐adjusted p = 0.005). AEC was highest in Group 1 (mean 522 cells/μL; FDR‐adjusted p = 0.039). Atopy rates differed in unadjusted comparisons (p = 0.0085) but did not survive FDR adjustment (FDR‐adjusted p = 0.102), and smoking status was not significant either before or after correction. Most other clinical variables—including asthma control, airflow limitation severity, IgE levels, FeNO, sputum neutrophils and eosinophils, and lung function—did not differ significantly across metabotypes (Figure 2). These results indicate that the metabotypes are primarily differentiated by age, sex, BMI, and peripheral eosinophilia, whereas asthma‐specific clinical features remain broadly comparable.

TABLE 1.

Demographic and clinical characteristics of the study population by cluster group. Metabolic features represent the relative abundance patterns of major metabolite classes in each group (↑: Increased, ↓: Decreased). Values are presented as mean ± standard deviation for continuous variables and the number of subjects for categorical variables.

Characteristics	Group 1 (n = 123)	Group 2 (n = 53)	Group 3 (n = 158)	Group 4 (n = 73)	The number of data by group	p	FDR‐adj p
METABOLIC	PC‐O↑	LPC↑, LPC‐O↑	SM↓	TG↑, DG↑	—	—	—
CLINICAL	—	—	—	—	—	—	—
Age (years) (min, max)	44.94 ± 14.68 (19, 81)	53.09 ± 14.44 (18, 79)	53.29 ± 16.2 (16, 85)	53.88 ± 14.63 (15, 81)	123/53/158/73	< 0.001	< 0.001
Age of symptom onset (years)	38.88 ± 14.73	43.59 ± 19.36	45.98 ± 16.5	45.94 ± 17.47	113/49/146/68	< 0.001	0.0074
Sex (female/male)	92/31	36/17	79/79	42/31	123/53/158/73	< 0.001	0.0056
BMI (kg/m2)	23.79 ± 3.649	24.68 ± 3.414	24.85 ± 3.792	26.14 ± 3.84	123/53/158/73	< 0.001	0.0053
Smoking status (current/ex/never)	13/28/82	5/10/38	21/56/81	11/23/38	123/53/158/72	0.0600	0.4798
Atopy a (n/y)	12/62	15/16	24/66	11/30	74/31/90/41	0.0085	0.1016
Allergic rhinitis (n/y)	5/66	1/19	7/68	4/35	71/20/75/39	0.8648	1.0000
Chronic rhinosinusitis (n/y)	7/2	5/0	11/0	7/1	9/5/11/8	0.3092	0.6747
Atopic dermatitis (n/y)	10/14	4/4	13/9	8/2	24/8/22/10	0.2149	0.5731
Diabetes mellitus (n/y)	9/0	5/0	10/3	5/2	9/5/13/7	0.2394	0.5745
Hypertension (n/y)	6/4	4/1	7/6	4/3	10/5/13/7	0.7862	1.0000
Gastroesophageal reflux disease (n/y)	5/4	4/1	12/1	6/1	9/5/13/7	0.2013	0.5731
Exacerbations (Ever) (none or not available/at least once)	79/38	28/24	109/44	43/29	117/52/153/72	0.0841	0.5048
Exacerbations (previous year) (none or not available/at least once)	18/29	7/17	22/33	14/19	47/24/55/33	0.7648	1.0000
ICS use (n/y)	52/32	14/18	62/43	22/22	84/32/105/44	0.2457	—
Asthma Control status (controlled/partly controlled/uncontrolled)	39/25/45	16/16/15	54/36/53	27/21/14	109/47/143/62	0.2494	—
ACT score	16.86 ± 4.932	17.15 ± 4.653	17.2 ± 4.628	18.08 ± 4.339	109/47/143/62	0.4838	0.6988
Airflow limitation severity (mild/moderate/severe)	53/36/30	22/18/10	60/60/29	37/22/13	119/50/149/72	0.5140	0.8812
Total IgE (IU/mL)	403.7 ± 666.5	304.5 ± 450.9	458.7 ± 756.6	440.8 ± 507.5	113/40/139/64	0.1667	0.4645
Absolute blood eosinophils (cells/μL)	522.2 ± 485.9	370.5 ± 429.9	362.2 ± 300.6	405.2 ± 361.2	123/53/155/72	0.0089	0.0387
Induced sputum neutrophils (%)	51.64 ± 32.28	48.56 ± 32.45	58.05 ± 32.99	53.99 ± 32.32	117/45/145/71	0.1841	0.4787
Induced sputum eosinophils (%)	10.83 ± 17.45	11.74 ± 17.89	7.967 ± 15.34	10.7 ± 17.57	86/35/118/52	0.7400	0.8488
FeNO (ppb)	58.41 ± 48.9	50.19 ± 34.91	57.64 ± 50.64	51.27 ± 31.39	69/28/91/35	0.9345	0.9591
Pre‐BD FEV1 predicted value (%)	74.43 ± 19.32	73.92 ± 18.89	74.19 ± 17.05	85.22 ± 62.44	123/53/156/72	0.3496	0.6253
Pre‐BD FEV1/FVC ratio	0.7357 ± 0.1382	0.7144 ± 0.09771	0.701 ± 0.1166	0.7205 ± 0.09696	119/50/149/72	0.1658	0.4645

Abbreviations: ACT, asthma control test; BD, bronchodilator; BMI, body mass index; FeNO, fractional exhaled nitric oxide; FEV1, forced expiratory volume in 1 s; FVC, forced vital capacity; ICS, inhaled corticosteroid. Note: Bold values denotes p or FDR‐adjusted p < 0.05.

^a Atopy was defined as a positive skin prick test to at least one common aeroallergen, or a clinical history of atopic dermatitis or allergic rhinitis.

FIGURE 2.

Distribution of clinical and demographic characteristics across the four metabotype clusters (G1–G4). Pairwise significance brackets within a panel are drawn only when the variable's omnibus test passed Benjamini–Hochberg multiple testing correction (FDR‐adjusted p < 0.05). Pairwise comparisons use Tukey HSD or Mann–Whitney U with Bonferroni correction for continuous variables; pairwise chi‐square tests with Bonferroni correction for categorical variables.

We then assessed whether these clinical associations persisted after joint adjustment for age, sex, BMI, log‐eosinophil count, smoking status, and ICS use, using a multinomial logistic regression with Group 3 as the reference (Table S5, Figure S13). The defining clinical axes of each non‐reference group remained after multivariable adjustment: younger age (adjusted OR 0.56, 95% CI: 0.42–0.73, p < 0.001) and female sex predominance (adjusted OR 0.36, 95% CI: 0.19–0.67, p = 0.001) for Group 1; lower log‐eosinophil count for Group 2 (adjusted OR 0.66, 95% CI: 0.50–0.87, p = 0.003); and higher BMI for Group 4 (adjusted OR 1.45, 95% CI: 1.09–1.93, p = 0.010). The unadjusted association between Group 1 and elevated eosinophils (OR 1.40, p = 0.021) attenuated to non‐significance after multivariable adjustment (adjusted OR 1.22, p = 0.174), indicating that the elevated eosinophilia of Group 1 in unadjusted comparisons largely reflected its younger age structure rather than an independent eosinophilic signal.

4. Discussion

This study identified four distinct asthma metabotypes by unsupervised clustering of plasma metabolomic profiles, each defined by a dominant lipid signature (Group 1, PC‐O; Group 2, LPC together with altered amino acid metabolism; Group 3, broadly reduced SM and PC; Group 4, markedly elevated TG and DG). These metabolomic signatures complement traditional inflammatory markers, such as blood AECs and FeNO, which have variable utility in predicting treatment response [11, 12]. Notably, Groups 3 and 4 shared a T2‐low clinical profile but were sharply distinguished by contrasting lipid patterns—SM depletion vs. neutral lipid accumulation—underscoring the metabolic heterogeneity within non‐T2 asthma. Conversely, PC‐O were elevated in Group 1 but selectively depleted in Group 4, suggesting that ether phospholipid homeostasis may be differentially regulated across asthma phenotypes. These findings underscore the complexity of asthma pathophysiology and support metabolomics‐driven precision medicine.

Although asthma is currently classified into phenotypes based on clinical features, inflammatory patterns, and triggering factors, variations in treatment response within the same phenotype and overlap between subtypes highlight the limitations of this approach. Endotype‐based classification has therefore emerged as a complementary framework focused on the biological pathways underlying disease mechanisms [13]. Earlier metabolomics studies, including those by Comhair et al. [7] and Reinke et al. [8] demonstrated asthma‐associated metabolic shifts but did not provide a comprehensive within‐asthma metabotype classification. Our analysis of plasma metabolites in a large cohort of Korean adults with asthma addresses this gap and suggests a pathway for integrating metabolomics with conventional inflammatory biomarkers to develop personalized treatment strategies.

A major finding of this study was the identification of Group 4 as a distinct obesity‐related metabotype. Patients in this group had the highest mean BMI and a metabolomics signature dominated by neutral lipids, with TG and DG showing strong class‐level over‐representation. The Group 4 signature was further characterized by a shift toward shorter total acyl carbon numbers for TG, PC, and DG, and toward longer chains for SM—a chain‐length pattern consistent with hepatic de novo lipogenesis, a metabolic state commonly linked to obesity and dysregulated lipid metabolism [14, 15, 16]. Low AEC and FeNO levels in this group are suggestive of non‐T2 inflammation. The BMI association of Group 4 remained robust after multivariable adjustment, and 161 of 166 unadjusted significant lipid metabolites retained significance after additional adjustment for BMI, age, and sex, indicating that the TG‐loaded lipid signature is not a passive consequence of higher adiposity but reflects a metabolically distinct subgroup.

Obesity is linked to non‐T2 asthma, which is characterized by lower T2 biomarkers and a predominance of neutrophilic or pauci‐granulocytic airway inflammation [17, 18]. Experimental studies suggest that obesity exacerbates airway inflammation through non‐T2 cytokines, including IL‐6, IL‐17, and TNF‐α [19], and that associated metabolic dysfunction, such as insulin resistance and lipid accumulation, may contribute to airway hyper responsiveness and steroid insensitivity. Patients with obesity and asthma may therefore exhibit distinct inflammatory mechanisms that reduce ICS effectiveness, highlighting the need to reconsider therapeutic approaches for this population [20]. In this context, metabolic‐targeted therapies such as GLP‐1 receptor agonists may represent a promising avenue [21, 22]. Beyond weight reduction, these agents may improve metabolic dysfunction and obesity‐related airway inflammation in non‐T2 asthma [23], although clinical validation is needed.

Group 1 emerged as an early‐onset metabotype characterized by elevated PC‐O. Patients in this cluster were the youngest, with the earliest symptom onset, marked female predominance, and the highest AEC in unadjusted comparisons. PC‐O, including plasmalogens, participates in cellular antioxidant defense and the resolution of inflammation [14]; their selective accumulation in Group 1 may reflect a protective lipid response or altered ether‐phospholipid remodeling associated with this early‐onset T2 phenotype [14]. Notably, the unadjusted eosinophilia of Group 1 attenuated to non‐significance after multivariable adjustment, indicating that the elevated eosinophil counts in this group largely reflect its younger, female‐predominant demographic structure rather than an independent eosinophilic signal. The reciprocal direction of PC‐O species–elevated in Group 1 but depleted in Group 4–suggests that ether phospholipid homeostasis may be differentially regulated between these metabotypes and warrants mechanistic investigation. From a therapeutic standpoint, Group 1 may benefit from established T2‐targeted biologic therapies (anti‐IL‐5/IL‐5R, anti‐IL‐4/IL‐13, and anti‐TSLP). However, the attenuation of eosinophilia after multivariate adjustment suggests that treatment selection should not rely on AEC but rather integrate multiple T2 biomarkers and clinical characteristics, including FeNO, atopy, exacerbation history, and corticosteroid responsiveness.

Group 3 displayed a distinct signature characterized by globally reduced SM and broadly lower PC, suggesting impaired membrane lipid homeostasis and potentially altered epithelial barrier function. Clinically, this group comprised middle‐aged, non‐obese patients with relatively low AEC, consistent with a mixed or T2‐low asthma phenotype. The depletion of structural membrane lipids may predispose these patients to airway fragility [24], reduced steroid responsiveness [25], and increased susceptibility to environmental triggers [26], suggesting that non‐T2‐targeted approaches such as LAMA or TSLP blockade warrant further investigation in this metabotype.

Group 2 represented a small but metabolically distinct cluster (n = 53) defined by elevated LPC and LPC‐O, together with altered amino acid metabolism—most prominently elevated glutamate and reduced glutamine. LPC species are bioactive lipids generated by phospholipase A2‐mediated hydrolysis of PCs and have been implicated in pro‐inflammatory signaling and oxidative stress in airway disease [27]. The concomitant rise in glutamate and fall in glutamine may reflect altered glutaminolysis or amino acid turnover, processes previously linked to immune cell activation [28] and asthma severity [29]. This dual lipid‐ and amino acid‐driven signature distinguishes Group 2 from purely T2‐high or T2‐low metabolic profiles, suggesting a metabolically intermediate subgroup whose biological underpinnings warrant further investigation. Therapeutically, the absence of a clear T2‐high signal suggests that conventional eosinophil‐targeted biologics may not be the optimal first‐line choice for most patients in this cluster; standard ICS‐based stepwise therapy with anti‐TSLP as a potential broader‐spectrum alternative may be more appropriate. This metabotype may also represent a clinically relevant “intermediate” subgroup frequently encountered in practice, in whom conventional T2 biomarkers incompletely explain symptom burden or treatment response.

Several limitations should be acknowledged. First, the absence of a healthy control group limited our ability to determine whether the identified signatures are specific to asthma pathobiology or partially reflect broader systemic metabolic states such as obesity or insulin resistance. Second, medication use may have influenced metabolite levels, as patients were on different treatments at sampling; future studies should include medication‐naïve patients or adjust for drug effects. Third, dietary habits, fasting status, and exact sampling time were not recorded, and the broad recruitment period (2005 to 2020) may have introduced variability in treatment background as clinical practice evolved, limitations common to plasma metabolomics studies in real‐world clinical cohorts. Fourth, the cross‐sectional design could not capture longitudinal changes in metabolite profiles; prospective studies are needed to assess the stability of these metabotypes and their predictive value. Fifth, the biological explanations proposed for each metabotype have not been confirmed by functional or experimental studies and should therefore be regarded as hypothesis‐generating rather than mechanistically established. Finally, the boundaries between metabotypes may reflect a continuum of metabolic states rather than strictly discrete biological entities, an inherent limitation of unsupervised clustering applied to continuous metabolomic data; whether these subgroups are truly distinct or represent gradual differences along a spectrum remains to be confirmed through longitudinal and functional studies.

5. Conclusion

This study demonstrates that plasma metabolomics can refine asthma classification, complement traditional inflammatory markers, and advance personalized treatment approaches. By identifying distinct metabolic signatures linked to lipid dysregulation, eosinophilic inflammation, and metabolic dysfunction, our findings underscore the value of integrating metabolomics into asthma management. Future research should validate these metabotypes in independent cohorts and explore targeted interventions that address specific metabolic pathways. Integrating metabolomics‐based classification with existing clinical frameworks may ultimately yield more precise, effective treatments, improving outcomes for patients with asthma.

Author Contributions

Woori Chae: methodology, software, formal analysis, investigation, writing – original draft, writing – review and editing, visualization. Jin An: formal analysis, writing – original draft, writing – review and editing. Chae Eun Lee: investigation. Seo‐Young Kim: data curation. Eunse Kim: data curation. Hyouk‐Soo Kwon: resources. Woo‐Jung Song: resources. You Sook Cho: resources. Joo‐Youn Cho: conceptualization, investigation, resources, writing – review and editing, supervision, project administration, funding acquisition. Tae‐Bum Kim: conceptualization, investigation, resources, data curation, writing – review and editing, supervision, project administration, funding acquisition.

Funding

This work was supported by Korea Health Industry Development Institute, RS‐2024‐00403700.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Workflow for metabotype clustering of patients with asthma using plasma metabolomics. Plasma samples from 407 adult patients with asthma in the COREA cohort were analyzed using targeted metabolomics with the Biocrates AbsoluteIDQ p400 HR kit and Q Exactive Plus Orbitrap mass spectrometer coupled with an Ultimate 3000 UPLC system. Data were preprocessed through concentration calculation, missing‐value imputation, and batch‐effect correction. Cluster analysis was performed using the k‐means algorithm to identify distinct metabolic metabotypes. Finally, key metabolites and clinical characteristics were interpreted to define important features of each cluster. Created in BioRender. https://BioRender.com/3rmttq8.

Figure S2: Metabolic profile distribution of the participants before and after QC‐based signal correction grouped by k‐means algorithm.

Figure S3: PCA projections of plasma metabolomic data across five clustering algorithms (KMeans, KMedoids, GMM, Spectral, Agglomerative) and five cluster numbers (k = 2–6). Each panel shows 407 patients with asthma colored by cluster assignment. Random seed was fixed at 24.

Figure S4: Cluster validity indices across cluster numbers (k = 2–6) for five clustering algorithms. Silhouette score, Calinski–Harabasz index, and Davies–Bouldin index are shown as mean ± SD across multiple random seeds; gap statistic (computed for k‐means only) is shown as value ± SE.

Figure S5: Consensus clustering results for k = 2–6 based on bootstrap resampling (B = 1000, 80% subsampling) using k‐means clustering. Heatmap shows consensus matrices with samples reordered by hierarchical clustering, where color intensity represents the proportion of bootstrap iterations in which each pair of samples was assigned to the same cluster (1.0 = always co‐clustered, 0.0 = never co‐clustered). The bottom‐right panel shows empirical cumulative distribution functions (CDFs) of consensus index values across all sample pairs for each k.

Figure S6: k‐means clustering stability across random seeds for k = 2–6. Seed‐consensus matrices (top five panels) show the proportion of seed pairs in which each pair of samples was assigned to the same cluster. Bottom three panels show partition agreement across seeds (adjusted Rand index, ARI), final inertia distribution, and the fraction of seed pairs producing identical partitions (ARI = 1), based on 100 random seeds per k.

Figure S7: Method‐vs‐method partition agreement. Heatmaps of the adjusted Rand index (ARI; 0 = chance agreement, 1 = identical partition) between cluster assignments produced by five clustering methods (k‐means, k‐medoids, GMM, spectral, agglomerative) at k = 3 and k = 4. Cell values show pairwise ARI. k‐means agrees substantially with GMM, spectral, and agglomerative clustering at both k values; k‐medoids produces partitions that diverge from the other four methods.

Figure S8: Distribution of plasma metabolites across the four metabotypes. (A, B) PCA scatter plot showing 407 asthma patients colored by metabotype assignment. (C) FDR‐adjusted p‐values from one‐way ANOVA or Kruskal–Wallis test (selected based on normality assessment) across all 281 metabolites.

Figure S9: Cluster‐discriminative metabolites across the four metabotypes. Heatmap of metabolites identified as discriminative for at least one metabotype, defined as those with absolute standardized effect size |Cohen's d| ≥ 0.8 and FDR‐adjusted p < 0.05 in one‐vs‐rest Welch's t‐tests on natural‐log‐transformed concentrations. For visual readability, the heatmap was capped at the top 20 metabolites per cluster ranked by |Cohen's d|; the complete set of cluster‐discriminant metabolites without this cap is provided in Table S3.

Figure S10: Lipid‐class enrichment per metabolite cluster. Heatmap of signed –log10 (FDR‐adjusted p) from Fisher's exact one‐sided over‐representation tests of each metabolite class within each cluster's discriminative metabolite set, versus its prevalence in the full 281‐metabolite panel. Stars indicate FDR‐adjusted p < 0.05 (*), < 0.01 (**), and < 0.001 (***).

Figure S11: Comparison of unadjusted vs. BMI/age/sex‐adjusted Group 4 effect sizes for 166 lipid metabolites. Colored circles, lipids retaining FDR‐adjusted p < 0.05 after adjustment; gray circles, lipids losing significance.

Figure S12: Chain‐length and saturation distributions for Group 4's cluster‐discriminative lipid metabolites. Within each panel, gray circles represent all class members in the metabolomics panel, and colored circles represent the metabolites belonging to Group 4's cluster‐discriminative set (|Cohen's d| ≥ 0.8 and FDR‐adjusted p < 0.05 in the one‐vs‐rest comparison).

Figure S13: Forest plot of unadjusted and adjusted odds ratios for each non‐reference metabotype cluster (G1, G2, G4) vs. the reference cluster (G3), from univariate and multivariable multinomial logistic regressions adjusting for age, sex, BMI, log‐transformed absolute eosinophil count, smoking status, and ICS use. Horizontal bars indicate 95% confidence intervals; the dashed line at OR = 1 represents the null effect, and the x‐axis is on a log scale.

Table S1: Fisher's exact enrichment of metabolite classes within each metabotype's cluster‐discriminative metabolite set.

Table S2: Per‐cluster retention of metabolite signals after multivariable clinical adjustment.

Table S3: Group 4 lipid metabolite effect sizes before and after adjustment for BMI, age, and sex.

Table S4: List of cluster‐discriminative metabolites for each metabotype.

Table S5: Unadjusted and mutivariable‐adjusted odds ratios for each clinical covariate by metabotype cluster.

Data Availability Statement

The data that support the findings of this study are openly available in Korea BioData Station at https://kbds.re.kr/, reference number KAP240302. All analysis scripts and the processed metabolomics concentration matrix are publicly available at GitHub (https://github.com/yunuschae/COREA_metabotype).