OP0277 RNA SEQUENCING AND MACHINE LEARNING TECHNIQUES PREDICT MAJOR ORGAN INVOLVEMENT IN PATIENTS WITH SYSTEMIC LUPUS ERYTHEMATOSUS. (June 2019)

• Record Type: Journal Article
• Authors: Filia, Anastasia; Bertsias, George; Panousis, Nikolaos; Nikolopoulos, Dionisis; Dermitzakis, Emmanouil; Boumpas, Dimitrios
• Is Part Of: Annals of the rheumatic diseases. Volume 78(2019)Supplement 2
• Page Start: 220
• Page End: 220
• Publication Date: 2019-06
• Subjects: Rheumatism -- Periodicals; 616.723005
• DOI: 10.1136/annrheumdis-2019-eular.4558
• Languages: English
• ISSNs: 0003-4967

Abstract

Abstract : Background: Both clinically and molecularly, Systemic Lupus Erythematosus (SLE) is a heterogeneous disease with non-synchronous multi-organ involvement of varying severity, and alternating periods of remission and flares. There is an unmet need for a blood-based- "liquid biopsy" to predict prognosis of the disease. Objectives: To detect the smallest set of genes predicting SLE organ involvement and disease activity using RNA-sequencing data derived from whole blood cells from 150 SLE patients and machine learning techniques. Methods: Disease activity was measured by the SLE disease activity index-2000 (SLEDAI-2K) and by organ involvement (treated as binary outcomes). SLEDAI-2K: mild-moderate disease (SLEDAI 0-8), severe (SLEDAI ≥8). Organ involvement: major (renal, heart, lung, central nervous system & SLEDAI-2K≥6) and minor (all others). The RNA-sequencing dataset was pre-processed to assemble 20, 368 genes and then split in training/validation data. Two feature selection steps (edgeR and recursive feature elimination) were used to remove noise and keep the smallest set of genes which best predicts each outcome. Different prediction models were fit to identify which one performs best using the gene signature selected in the previous step. Results: Two gene signatures were kept after feature selection to predict each of the two outcomes (25 genes for organ involvement; 50 genes for SLEDAI-2K). Organ involvement was predicted with high accuracy (accuracy=0.89, sensitivity=0.89, specificity=0.88 in the validation data) using the elastic net generalised linear model. Among the 25 best predictors were MPO, ITGA3 and CD38 . SLEDAI-2K could not be predicted with high accuracy (accuracy 0.75, sensitivity=0.79, specificity=0.67) using the neural network model. Performance was still the same even when 1648 genes (after first feature selection step) were used as predictors of SLEDAI-2K. The performance of these models will also be tested in an independent test dataset once available (currently undergoing sequencing). Conclusion: The model predicting organ involvement performed better compared to the model predicting SLEDAI-2K. This could be attributed to the fact that certain disease manifestations are not currently included in SLEDAI-2K. Further analysis and functional laboratory experiments of those genes will help to identify biomarkers for more accurate assessment of disease activity and prognosis in the clinic. Acknowledgement: This work was supported by FOREUM, SYSCID and ERC-Advanced Grant. Disclosure of Interests: None declared