Machine Learning-Based Identification of Novel Exosome-Derived Metabolic Biomarkers for the Diagnosis of Systemic Lupus Erythematosus and Differentiation of Renal Involvement

Objective: This study aims to investigate the exosome-derived metabolomics profiles in systemic lupus erythematosus (SLE), identify differential metabolites, and analyze their potential as diagnostic markers for SLE and lupus nephritis (LN).

Methods: Totally, 91 participants were enrolled between February 2023 and January 2024 including 58 SLE patients [30 with nonrenal-SLE and 28 with Lupus nephritis (LN)] and 33 healthy controls (HC). Ultracentrifugation was used to isolate serum exosomes, which were analyzed for their metabolic profiles using liquid chromatography-tandem mass spectrometry (LC-MS/MS). Endogenous metabolites were identified via public metabolite databases. Random Forest, Lasso regression and Support Vector Machine Recursive Feature Elimination (SVM-RFE) algorithms were employed to screen key metabolites, and a prediction model was constructed for SLE diagnosis and LN discrimination. ROC curves were constructed to determine the potential of these differential exosome-derived metabolites for the diagnosis of SLE. Furthermore, Spearman's correlation was employed to evaluate the potential links between exosome-derived metabolites and the clinical parameters which reflect disease progression.

Results: A total of 586 endogenous serum exosome-derived metabolites showed differential expression, with 225 exosome-derived metabolites significantly upregulated, 88 downregulated and 273 exhibiting no notable changes in the HC and SLE groups. Machine learning algorithms revealed three differential metabolites: Pro-Asn-Gln-Met-Ser, C24:1 sphingolipid, and protoporphyrin IX, which exhibited AUC values of 0.998, 0.992 and 0.969 respectively, for distinguishing between the SLE and HC groups, with a combined AUC of 1.0. In distinguishing between the LN and SLE groups, the AUC values for these metabolites were 0.920, 0.893 and 0.865, respectively, with a combined AUC of 0.931, demonstrating excellent diagnostic performance. Spearman correlation analysis revealed that Pro-Asn-Gln-Met-Ser and protoporphyrin IX were positively correlated with the SLE Disease Activity Index (SLEDAI) scores, urinary protein/creatinine ratio (ACR) and urinary protein levels, while C24:1 sphingolipid exhibited a negative correlation.

Conclusions: This study provides the first comprehensive characterization of the exosome-derived metabolites in SLE and established a promising prediction model for SLE and LN discrimination. The correlation between exosome-derived metabolites and key clinical parameters strongly indicated their potential role in SLE pathological progression.