Background/Purpose: Objectives: Systemic lupus erythematosus (SLE) is an autoimmune disease marked by unpredictable flares. However, no flare risk prediction model has been developed for Asian populations. This study aimed to construct a causal proteomics-defined risk prediction model to predict subsequent flares.

Methods: Methods: 139 SLE patients who were followed up every 12 weeks, were recruited for the study. After 48 weeks, they were divided into flare (n = 53) and non-flare groups (n = 86) to assess the significance of clinical predictors using logistic regression (LR) and random forests (RFs). Plasma data independent acquisition (DIA) proteomics was performed at the baseline to identify biomarkers of flare and to conduct biological pathway analyses. Protein quantitative trait loci (pQTLs) from the UK Biobank study were used to perform phenome-wide Mendelian randomization (PheWAS) analysis to find causal effects between identified flare circulating protein on SLE and clinical predictors using 220 disease outcomes in BioBank Japan (n= 179,000). Machine learning methods were used to fit the models by causal proteomics results and compared them with a clinical risk model.

Results: Results: The plasma proteome revealed 102 proteins including 73 upregulated proteins and 23 downregulated proteins that were significantly different between flare and non-flare patients. Non-downregulated proteins including SAA1, B4GALT5, GIT2, NAA15 and RPIA were positively associated with SLEDAI and 1-year flare. According to the Reactome pathway, upregulated flare genes were mainly enriched in B cell Receptor(BCR), response to elevated platelet cytosolic Ca2+ and platelet degranulation pathways. Our PheWAS evidence suggested SAA1 had causal effects on flare predictors including chronic glomerulonephritis (OR = 0.738, 95% CI: 0.592-0.920, P = 0.007). UTP, NLR, PLR, rRNP, lupusQoL, SLEDAI-2K, and baseline glucocorticoids were selected as variables for the clinical prediction model. Cross-model analysis showed a combinatorial biomarker more accurately distinguished patients with flare from non-flare individuals with an area under the curve of 0.767 (AUC = 0.767) than only the proteins model (AUC =  0.744) and clinical model (AUC =  0.628).

Conclusion: Conclusions: Our findings indicated that meaningful biomarkers associated with SLEDAI, as determined in PheWAS, predicted flares more accurately than only clinical profiles. SAA1 could be prioritized for rapid discrimination of flares.

« Back to ACR Convergence 2024

ACR Meeting Abstracts - https://acrabstracts.org/abstract/causal-proteomics-assisted-machine-learning-model-enhances-flare-risk-prediction-in-systemic-lupus-erythematosus/