Citrullination is a crucial autoantigen modification in autoimmune diseases such as rheumatoid arthritis. However, accurate citrullinome profiling remains limited by insufficient detection methods, a lack of computational tools, and the large sample input required by current techniques. In this study, we developed a low-input deep learning quantitative analysis platform named Iseq-Cit, which achieves enrichment-free, high-throughput citrullinome profiling while requiring less than 1% of the sample input needed for conventional methods. Utilizing this platform, we comprehensively profiled a longitudinal cohort of individuals at risk for RA and patients with confirmed RA, discovering that plasma citrullinome profiles closely correlate with RA development and severity. Furthermore, by integrating clinical indicators with omics data, we constructed machine learning models capable of predicting specific drug treatment responses with high accuracy. Simultaneously, we trained a deep learning model based on a bidirectional gated recurrent unit (BiGRU), which successfully predicted the RA-sera reactivity of peptides with 84.2% accuracy, yielding 19 promising novel candidate autoantigens for clinical diagnosis. This work provides novel strategies for precise RA treatment stratification and autoantigen discovery.

    This study included 711 plasma samples, encompassing healthy controls, RA at-risk and rheumatoid arthritis. The Iseq-Cit technology generated hyper-citrullinated peptides as internal standards (IS) via in vitro enzymatic catalysis of multi-organ proteins using PAD2 and PAD4. Subsequently, tandem mass tags were utilized for isotope labeling, and the internal standards and samples were mixed at a 2:1:1 ratio; this approach of amplifying the sample's citrullination signals was used to map the plasma citrullinome of RA. Based on the citrullinomic data, feature selection was performed using methods such as LASSO and random forest, and KNN algorithm was employed to construct a classifier for the drug treatment response prediction model. For autoantigen prediction, a bidirectional gated recurrent unit (BiGRU) deep learning model with a self-attention mechanism was trained based on the physicochemical features of 67,399 negative and 8,816 positive peptides, followed by antigen screening and validation via ELISA and evaluations of T cell activation levels.

    Using extremely low sample input, the Iseq-Cit platform identified 1,536 plasma and 1,026 synovial citrullinated peptides with high confidence. Omics analysis indicated that abnormal citrullination is activated before the onset of RA , and the abundance of multiple modified peptides showed a significant negative correlation with disease severity indicators, such as DAS28 and swollen joint count. The KNN predictive model, which integrated clinical features and citrullination data, demonstrated excellent performance in independent external validation, achieving AUROCs of 88.3% and 79.4% for predicting therapeutic responses to MTX combined with LEF or HCQ, respectively. Furthermore, we established a deep learning model achieving an accuracy of 92.2% in the task of distinguishing RA-sera reactivity peptide. Upon external validation via ELISA, the accuracy reached 84.2%, successfully identifying 19  autoantigens with high disease specificity. T cell functional assays further confirmed that specific peptides could significantly increase the percentage of CD154⁺CD4⁺ T cells and promote IFN-γ release, demonstrating potent immunogenicity.

This study successfully established an ultra-low input, enrichment-free Iseq-Cit mass spectrometry platform, comprehensively mapping the systemic citrullinome landscape of healthy individuals, RA at-risk, and diagnosed patients. The research revealed the dynamic evolutionary patterns of the citrullination modification network during the early stages of RA pathogenesis, providing valuable molecular evidence for preclinical disease warning. By deeply integrating characteristic omics data with clinical indicators, an RA drug response prediction model was established, significantly advancing individualized and precise treatment for RA. Simultaneously, utilizing a deep learning model combined with an antigen validation system enabled the highly efficient and accurate screening and identification of novel pathogenic autoantigens. This work paves a new path for antigen discovery and stratified diagnosis and treatment.