A quarter of the world"s population suffers from chronic inflammatory skin diseases (ISDs). These diseases encompass a broad spectrum of heterogenous entities and a wide range of different diagnoses. As a consequence, ISDs are particularly challenging to diagnose accurately and timely, which is a prerequisite for the effective treatment of the underlying disease.

Omic technologies in combination with artificial intelligence hold the promise to contribute substantially to clinical decision making and patient care. Here, we performed mass spectrometry-based proteomics with the latest generation mass spectrometer (Orbitrap Astral) from Thermo Fisher to generate a comprehensive proteomic atlas of all major inflammatory skin diseases. For this, we collected retrospective FFPE skin tissue biopsies and associated clinical metadata from 459 patients, so far. These FFPE tissue sections were processed for MS-based proteomic analysis using a semi-automated (Opentrons) workflow and integrated normalization strategies, with the ultimate goal in mind to enable this workflow at other locations in the near future. All samples were rapidly measured using the standardized 21 min gradient (60SPD) on the Evosep One connected to the Orbitrap Astral (Thermo) using an 8cm PepSep column.

This robust setup yielded 10,364 proteins across all diseases measured, spanning five orders of magnitude and contained many low-abundant cytokines. With our comprehensive proteomic dataset and an independent additional cohort at hand we assessed the capability of our atlas to improve clinical decision making in "Erythroderma" (n = 115). Erythroderma is a potentially life-threatening condition defined by redness that affects more than 90% of the skin and is the result of an underlying inflammatory skin condition, including psoriasis, atopic dermatitis, pityriasis rubra pilaris, cutaneous T-cell lymphoma and drug reactions. The diagnosis of the underlying cause usually takes many weeks to months, despite the need of rapid and disease specific treatment initiation. Through integration of machine learning algorithms, our proteomic dataset effectively distinguished between the erythrodermic diseases with high accuracy and an AUC of up to 0.92, using up to 50 proteins. Feature selection was performed across these two independent cohorts by bootstrapping and permutation testing with 5-fold cross-validation. In addition, we detected novel and disease specific biology, including the induction of the IL17C/IL17RE Axis in pityriasis rubra pilaris. This study successfully highlights the potential of proteomics for AI-driven disease stratification in cutaneous inflammatory disorders and clinical decision making in general.