Integrative Pharmacovigilance and AI-Based Framework Uncovers Potential Drug Triggers in Juvenile Idiopathic Arthritis

Luo Qiang¹Yuxiao Chen¹Dawei Liu¹Xinlin Wu¹Jingyi Zhao²Xuemei Tang^1*

• ¹Children‘s Hospital of Chongqing Medical University, Chongqing, China
• ²Chengde Medical University, Chengde, China

Background Management of juvenile idiopathic arthritis (JIA) relies heavily on long-term pharmacotherapy, yet an increasing number of case reports suggest that some drugs may themselves precipitate or worsen the disease. But systematic methods for detecting these safety signals in paediatric cohorts are still lacking. Methods We screened 10,012,438 reports from the FAERS database using four disproportionality algorithms (ROR, PRR, EBGM, and BCPNN) to identify potential drug and JIA associations. Three complementary machine learning models were developed, including DMPNN, GCN, and SVM, trained on molecular descriptors, chemical fingerprints, and structural graphs to stratify high-risk compounds. Toxicogenomic profiles were generated using ProTox-3.0, and drug–disease target overlap and pathway enrichment were assessed using the CTD and GeneCards databases. External validation relied on our own newly generated transcriptomic data: (i) our newly generated bulk RNA-seq dataset from 47 individuals (39 JIA patients and 8 controls) and (ii) a multi-centre single-cell RNA-seq compendium that combined 21 in-house PBMC profiles obtained at four Chinese paediatric hospitals with 9 publicly available systemic juvenile idiopathic arthritis (sJIA) samples. Two of the in-house sJIA patients were sampled longitudinally, before and one month after IL-6-receptor-inhibitor therapy permitting assessment of treatment-induced transcriptomic shifts. Drug-signature activity was quantified with single-sample GSEA for the bulk data and AddModuleScore for the single-cell data. Results We identified drugs with consistent positive signals across all four FAERS-based disproportionality algorithms. Machine learning models (DMPNN, GCN, SVM) independently confirmed 23 high-risk compounds, with 22 overlapping across all models and predicted risk scores >0.60. Among these, lansoprazole and aripiprazole showed strong signals in both pharmacovigilance and DMPNN predictions. Further toxicogenomic analysis revealed immune toxicity patterns overlapping with JIA-related gene targets and pathways. Notably, bulk RNA-seq and single-cell RNA-seq validation demonstrated that lansoprazole signatures were significantly enriched in monocyte from sJIA patients. This multi-level convergence supports the hypothesis that certain non-antirheumatic drugs may aggravate JIA-like inflammation, particularly within the systemic subtype. Conclusions In this study, we identify lansoprazole as a likely instigator of systemic juvenile idiopathic arthritis, underscoring that proton-pump inhibitors should be used judiciously in children at autoimmune risk and providing a generalisable playbook for rare-disease pharmacovigilance.