Artificial Intelligence‐Driven Proteomics Identifies Plasma Protein Signatures for Diagnosis and Stratification of Behçet's Disease

The diagnosis of Behçet's disease (BD) predominantly relies on clinical symptoms, indicating an urgent for identifying potential biomarkers for early diagnosis and disease stratification. We employed an in-depth proteomics platform based on data-independent acquisition mass spectrometry (DIA-MS) and customizable antibody microarray technology, combined with machine learning methods. By analyzing the proteomic data in the training cohort, we trained an XGBoost machine learning model, and validated the model in an independent cohort. The model displayed a favorable performance in BD diagnosis and stratification. In the training set, the area under the curve (AUC) of the diagnostic model was 0.984 with an accuracy of 0.935. In the validation set, the AUC was 0.967 with an accuracy of 0.871. The AUCs for differentiating different severity BD groups ranged from 0.897 to 0.986 in the training set, and from 0.718 to 0.960 in the validation set. Functional analysis indicated that processes such as defense response, protein activation cascade, and complement activation were related to disease severity. Complement C4B was crucial in the protein-protein interaction network. This study is the first to construct an artificial intelligence-based BD diagnosis and stratification model, providing potential biomarkers and new strategies for precise diagnosis and treatment of BD. The diagnosis of BD predominantly relies on clinical symptoms, indicating an urgent for identifying potential biomarkers for early diagnosis and disease stratification. We employed an in-depth proteomics platform based on data-independent acquisition mass spectrometry (DIA-MS) and customizable antibody microarray technology, combined with machine learning methods. By analyzing the proteomic data in the training cohort, we trained an XGBoost machine learning model, and validated the model in an independent cohort. The model displayed a favorable performance in BD diagnosis and stratification. In the training set, the area under the curve (AUC) of the diagnostic model was 0.984 with an accuracy of 0.935. In the validation set, the AUC was 0.967 with an accuracy of 0.871. The AUCs for differentiating different severity BD groups ranged from 0.897 to 0.986 in the training set, and from 0.718 to 0.960 in the validation set. Functional analysis indicated that processes such as defense response, protein activation cascade, and complement activation were related to disease severity. Complement C4B was crucial in the protein-protein interaction network. This study is the first to construct an artificial intelligence-based BD diagnosis and stratification model, providing potential biomarkers and new strategies for precise diagnosis and treatment of BD.