Background/Purpose: To refine clinical subtypes of systemic lupus erythematosus (SLE) using unsupervised machine learning, and to elucidate the molecular basis of disease heterogeneity through integrated immune, proteomic, and transcriptomic profiling.

Methods: We enrolled 1,065 SLE patients hospitalized at Peking University People’s Hospital between 2012–2024. Unsupervised hierarchical clustering was applied to integrated clinical and laboratory variables. Peripheral blood lymphocyte subsets were assessed to characterize immunologic divergence across clusters. One-year prospective follow-up was conducted to evaluate therapeutic responses. Multi-omics profiling—including serum proteomics (LC-MS/MS) and transcriptomics (bulk RNA-seq) of peripheral blood mononuclear cells—was performed in subtype-stratified patients and age-/sex-matched healthy controls. Differential expression and enrichment analyses were used to define cluster-specific molecular signatures.

Results: Unsupervised machine learning identified SLE subtypes: Cluster 1 exhibited nephritis, high disease activity, and anti-Sm positivity; Cluster 2 had skin/mucosal involvement with low inflammation; Cluster 3 was defined by neuropsychiatric manifestations and anti-U1RNP/SSA positivity; Cluster 4 featured anemia, leukopenia, and thrombocytopenia (85.9%), with Coombs positivity and antiphospholipid antibody enrichment; Cluster 5 involved serosal/vascular complications with older age and elevated LDH. Immune cell profiling revealed distinct T, B, and NK subset distributions. Rituximab was most effective in Cluster 5 (β = 0.286, p < 0.001), while cyclophosphamide benefited Cluster 3 (β = 0.284, p < 0.001). Cluster 2 showed the highest LLDAS (46.4%) and steroid tapering; Cluster 4 had the highest SLEDAI < 4 rate (84.1%). In addition, multi-omics analyses identified key cluster-enriched molecules: APOA4 and LAMC1 in Cluster 1; CXCL7 and CCL18 in Cluster 2 ; S100A7 and ACTN4 in Cluster 3 ; IGHG1 and SERPINA1 in Cluster 4; THBS1 and LOX in Cluster 5. Transcriptomics revealed shared upregulation of interferon-inducible genes (e.g., IFI44, IFITM1) and cluster-specific cytokines (e.g., WNT5B, KRT5, BDNF), highlighting distinct pathogenic pathways that parallel clinical phenotypes.

Conclusion: Machine learning–guided subtyping uncovers clinically and molecularly distinct SLE subgroups. These findings support biologically informed stratification and might lay the groundwork for precision therapy in SLE.

Hierarchical clustering tree and cluster reduction factor maps of systemic lupus erythematosus (SLE)

« Back to ACR Convergence 2025

ACR Meeting Abstracts - https://acrabstracts.org/abstract/machine-learning-defined-subtypes-of-systemic-lupus-erythematosus-identify-distinct-immunologic-and-molecular-signatures/