Abstract: BACKGROUND: Ferroptosis-related genes have been found to play an important role in the pathogenesis of rheumatoid arthritis. However, there is currently a lack of immune expression of ferroptosis-related signature genes in rheumatoid arthritis and the construction of competing endogenous RNA (CeRNA) interaction networks. Machine learning, as a powerful signature gene selection algorithm based on bioinformatics, can more accurately identify ferroptosis-related signature genes that dominate the pathogenesis of rheumatoid arthritis.
OBJECTIVE: To screen ferroptosis-related signature genes in rheumatoid arthritis using bioinformatics and machine learning methods, and to analyze the correlation between ferroptosis-related signature genes and immune infiltration and the construction of CeRNA network of ferroptosis-related signature genes.
METHODS: Rheumatoid arthritis-related microarrays were obtained from the GEO database, and ferroptosis-related genes and their differential gene expression were extracted using R language. The differentially expressed genes were screened using machine learning methods. The LASSO regression and SVM-RFE methods were used for signature gene screening, and the genes filtered by both were re-intersected to finally obtain the signature genes in rheumatoid arthritis. Receiver operating characteristic curves were used to assess the accuracy of the screened signature genes for disease diagnosis. Immune infiltration of rheumatoid arthritis and normal synovial tissues was analyzed using the CIBERSORT algorithm, and the correlation between the signature genes and immune cells was analyzed. Finally, the CeRNA network of ferroptosis-related signature genes for rheumatoid arthritis was constructed and the disease signature genes were validated. 
RESULTS AND CONCLUSION: A total of 150 ferroptosis-related genes in rheumatoid arthritis were obtained, including 55 up-regulated genes and 95 down-regulated genes. GO and KEGG enrichment analyses identified 18 GO significantly correlated entries and 30 KEGG entries respectively, mainly involving metal ion homeostasis, ferric ion homeostasis and oxidative stress response. Machine learning analysis finally identified disease signature genes GABARAPL1 and SAT1. GSEA analysis found that adipocytokine signaling pathway, drug metabolism cytochrome P450, fatty acid metabolism, PPAR signaling pathway, tyrosine metabolism were mainly concentrated when GABARAPL1 was highly expressed, and chemokine signaling pathway, intestinal immune network on IGA production were mainly concentrated when SAT1 was highly expressed. Immune infiltration analysis found that nine immune cells were significantly different in rheumatoid arthritis and normal synovial tissues, in which plasma cells, T-cell CD8, and T-cell follicular helper were highly expressed and the rest were lowly expressed in the disease group. Single gene and immune cell correlation analysis found that GABARAPL1 was positively correlated with dendritic resting cells, activated NK cells, and macrophage M1, with the most significant correlation with dendritic resting cells, while SAT1 was positively correlated with T cell CD4 and γδ T cells and negatively correlated with NK resting cells. GSVA analysis found that SAT1 was upregulated in ascorbic acid and aldehyde metabolism, while downregulated in B-cell receptor signaling pathway, Toll-like receptor signaling pathway, T-cell receptor signaling pathway, and natural killer cell-mediated cytotoxicity. GABARAPL1 showed a down-regulation trend in PPAR signaling pathway, metabolism of nicotinate and nicotinamide, tryptophan metabolism, fatty acid metabolism, and steroid biosynthesis. Sixty long non-code RNAs may play a key role in the development of rheumatoid arthritis. To conclude, the occurrence of rheumatoid arthritis is significantly correlated with the abnormal expression of rheumatoid arthritis-induced ferroptosis-related signature genes, and the signature genes induce disease development via relevant signaling pathways. By analyzing rheumatoid arthritis-related long non-code RNAs-mediated ceRNA networks, potential therapeutic targets and signaling pathways can be identified to further elucidate its pathogenesis and provide a reference basis for subsequent experimental studies.

Key words: rheumatoid arthritis, bioinformatics, machine learning, ferroptosis, CeRNA network construction