Accurate diagnosis of atopic dermatitis by applying random forest and neural networks with transcriptomic data

Abstract

Atopic dermatitis (AD) is one of the most common inflammatory skin diseases. But the great heterogeneity of AD makes it difficult to design an accurate diagnostic pipeline based on traditional diagnostic methods. In other words, the AD diagnosis has suffered from an inaccurate bottleneck. Thus, it is necessary to develop a novel and accurate diagnostic model to supplement existing methods. The recent development of advanced gene sequencing technologies enables potential in accurate AD diagnosis. Inspired by this, we developed an accurate AD diagnosis based on transcriptomic data in skin tissue. Using these data of 149 subjects, including AD patients and healthy controls, from Gene Expression Omnibus (GEO) database, we screened differentially expressed genes (DEGs) of AD and identified six critical genes (PPP4R1, SERPINB4, S100A7, S100A9, BTC, and GALNT6) by random forest classifier. In a follow-up study of these genes, we constructed a neural network model (average AUC=0.943) to automatically distinguish subjects with AD from healthy controls. Among these critical genes, we found that PPP4R1 and GALNT6 had never been reported to be associated with AD. Although further replications in other cohorts are needed, our findings suggest that these genes may be developed into useful biomarkers of AD diagnosis and may provide invaluable clues or perspectives for further researches on the pathogenesis of AD.