Machine learning applied to atopic dermatitis transcriptome reveals distinct therapy-dependent modification of the keratinocyte immunophenotype.

BACKGROUND Atopic dermatitis (AD) arises from a complex interaction between an impaired epidermal barrier, environmental exposures, and the infiltration of Th1/Th2/Th17/Th22 T cells. Transcriptomic analysis has advanced understanding of gene expression in cells and tissues. However, molecular quantitation of cytokine transcripts does not predict the importance of a specific pathway in AD or cellular responses to different inflammatory stimuli. OBJECTIVE To understand changes in keratinocyte transcriptomic programmes in human cutaneous disease during development of inflammation and in response to treatment. METHODS We performed in silico deconvolution of the whole-skin transcriptome. Using co-expression clustering and machine learning tools, we resolved the gene expression of bulk skin (n=7 datasets, n=406 samples), firstly, into keratinocyte phenotypes identified by unsupervised clustering and, secondly, into 19 cutaneous cell signatures of purified populations from publicly available datasets. RESULTS We identify three unique transcriptomic programmes in keratinocytes, KC1, KC2, KC17, characteristic to immune signalling from disease-associated helper T cells. We cross-validate those signatures across different skin inflammatory conditions and disease stages and demonstrate that the keratinocyte response during treatment is therapy dependent. Broad spectrum treatment with ciclosporin ameliorated the KC17 response in AD lesions to a non-lesional immunophenotype, without altering KC2. Conversely, the specific anti-Th2 therapy, dupilumab, reversed the KC2 immunophenotype. CONCLUSION Our analysis of transcriptomic signatures in cutaneous disease biopsies reveals the effect of keratinocyte programming in skin inflammation and suggests that the perturbation of a single axis of immune signal alone may be insufficient to resolve keratinocyte immunophenotype abnormalities.