Atopic dermatitis (AD) is a frequent and heterogeneous inflammatory skin disease, for which personalized medicine remains a challenge. High-throughput approaches have improved understanding of the complex pathophysiology of AD. However, a purely data-driven AD classification is still lacking.
To address this question, we applied an original unsupervised approach on the largest available transcriptome dataset of AD lesional (n = 82) and healthy (n = 213) skin biopsies.
Taking into account pathological and physiological state, a variance-based filtering revealed 222 AD-specific hyper-variable genes that efficiently classified the AD samples into 4 clusters that turned out to be clinically and biologically distinct. Comparison of gene expressions between clusters identified 3 sets of upregulated genes used to derive metagenes (MGs): MG-I (19 genes) was associated with IL-1 family signaling (including IL-36A and 36G) and skin remodeling, MG-II (23 genes) with negative immune regulation (including IL-34 and 37) and skin architecture, and MG-III (17 genes) with B lymphocyte immunity. Sample clusters differed in terms of disease severity (p = .02) and S. aureus (SA) colonization (p = .02). Cluster 1 contained the most severe AD, highest SA colonization, and overexpressed MG-I. Cluster 2 was characterized by less severe AD, low SA colonization, and high MG-II expression. Cluster 3 included mild AD, mild SA colonization, and mild expression of all MGs. Cluster 4 had the same clinical features as cluster 3 but had hyper-expression of MG-III. Last, we successfully validated our method and results in an independent cohort.
Our study revealed unrecognized AD endotypes with specific underlying biological pathways, highlighting novel pathophysiological mechanisms. These data could provide new insights into personalized treatment strategies.