The clinical phenotypes included within the broad “umbrella” definition of asthma are well known to be diverse (1, 2). Gene expression and genetic variation studies have confirmed the observed clinical heterogeneity by indicating that asthma is a polygenic disease with multiple molecular roots (3, 4). Yet refining these differences to the final development of distinct, universally accepted phenotypes to guide effective treatment choices has proven a challenge. Recently, the application of bioinformatic clustering techniques of clinical and inflammatory variables has identified modestly consistent phenotypes, but without the molecular characteristics essential to better understand underlying pathogenesis and best treatment approaches (5–7). Cross-sectional measurement of airway epithelial cell gene expression has demonstrated that approximately 50% of mild, corticosteroid-naive patients with asthma have up-regulation of a type-2 inflammation gene expression signature. These “type-2–high” patients with mild asthma demonstrated a tendency for atopy, eosinophilic airway inflammation, and a positive response to inhaled corticosteroids (8, 9). In a study from the Severe Asthma Research Program (SARP), clustering of patients with severe disease using only airway epithelial cell gene expression uncovered five subtypes. Highlighting the complexity of severe asthma, each of the five subtypes had distinct clinical characteristics and gene expression profiles that included type-2 genes but also included additional gene clusters representative of multiple diverse biological functions (4). In the current issue of the Journal, Kuo and colleagues (pp. 443–455) report gene expression findings from 91 bronchial biopsies and 99 epithelial brushings of 107 participants with moderate to severe asthma enrolled in the U-BIOPRED (Unbiased Biomarkers in Prediction of Respiratory Disease Outcomes) cohort (10). Eighty-three of these biopsy and brushing samples were matched, the largest matched tissue/epithelial sample to date. The cohort was enriched for severe asthma (z63%), with 27% using daily oral corticosteroids and a collective FEV1 of z75% predicted. In comparison, 39% of the SARP cohort was diagnosed as having “severe” asthma, with a collective FEV1 of z72% predicted. Kuo and colleagues used a bioinformatic method known as geneset variation analysis (GSVA)(10). In GSVA, the user inputs various lists of genes that they determine to be of particular interest. Here the authors chose 42 manually curated lists derived from a diverse array of prior gene expression experiments. Seven of the 42 lists came from asthma-specific studies, including in vitro cell studies, mouse and primate models, and the study of airway epithelial cells that initially defined type-2–high asthma (8, 9). For each sample’s gene expression profile, an enrichment score is calculated for each of the 42 gene lists. This is done by measuring enrichment for up-or down-regulation of genes within a gene list compared with all genes outside that set. This step transforms the data from a gene by sample matrix to a gene-set by sample matrix, where each data point of the gene-set by sample matrix is an enrichment score.

First using the bronchial biopsy gene expression profiles, unsupervised clustering of the gene-set by sample matrix identified two asthma subgroups: cluster A and cluster non-A. A phenotype classifier and feature selection tool then showed that only 9 of the 42 gene lists were needed to predict cluster A versus non-A with high accuracy. Subsequent application of these nine GVSA signatures to the 99 lung epithelial brushing samples ultimately resulted in the major finding of the …