Unsupervised machine learning identifies distinct SLE patient endotypes with differential response to belimumab

Abstract

Objective: To determine SLE endotypes according to B cell immunophenotyping and serological profile and assess endotypes' response to belimumab.

Methods: We analysed data from 796 patients with SLE from the phase III trial BLISS-SC. Unsupervised machine learning employing factor analysis of mixed data (FAMD) and subsequent clustering determined endotypes based on B cell immunophenotyping and serological profiles. Cox regression was used to assess belimumab efficacy on inducing lupus low disease activity state (LLDAS) and definitions of remission in SLE (DORIS) remission within clusters.

Results: Three endotypes were determined. Compared with each other, cluster 1 (n = 191) displayed higher proportions of CD19+CD24b+CD27+ regulatory B cells (mean ± SD: 35.9%±12.6%), CD19+CD20+CD27+ bulk memory B cells (32.2%±9.9%), CD19+CD20+CD69+ activated B cells (0.2%±0.3%), CD19+CD20-CD138+ long-lived plasma cells (0.7%±1.0%) and CD19+CD38b+CD27b+ SLE-associated plasma cells (6.6%±7.0%). Cluster 2 (n = 366) displayed higher proportions of CD19+CD24bbrightCD38bbrightCD27- transitional B cells (6.3%±9.0%) and CD19+CD20+CD27- naïve B cells (85.4%±7.2%), and lower proportions of CD19+CD20-CD138+ peripheral long-lived plasma cells (0.2%±0.3%) and CD19+CD38b+CD27b+ SLE-associated plasma cells (1.6%±2.0%). Cluster 3 (n = 239) displayed a higher proportion of CD19+CD20+CD138+ short-lived plasma cells (0.1%±0.1%) and higher serological activity, being enriched in anti-double stranded(ds)DNA, anti-ENAs, antiphospholipid antibodies and hypocomplementemia. Use of belimumab was superior to placebo in inducing sustained LLDAS [hazard ratio (HR): 2.22; 95% CI: 1.18-4.17; P = 0.014] and DORIS remission (HR: 3.45; 95% CI: 1.2-9.94; P = 0.022) in cluster 2.

Conclusion: Three distinct SLE endotypes were identified based on B cell immunophenotyping and serological profiles, showing differential benefit from belimumab therapy.

Keywords: B cells; SLE; belimumab; machine learning.

Figure 1.

t-SNE plot of patient clusters. T-distributed stochastic neighbour embedding (t-SNE) plot depicting the three clusters of patients with SLE identified by an unsupervised analysis using B cell subsets and serology

Figure 2.

B cell maturation pathway and cluster-specific differences. Schematic representation of the B cell maturation pathway, overlaid with the relative percentages of B cell subsets across the three identified patient clusters. The schematic illustrates key maturation stages of B cells, from transitional and naïve B cells to plasmablasts and long-lived plasma cells, highlighting differences in subset abundance across clusters

Figure 3.

Disease activity and organ damage at baseline by clusters. Distribution of scored (A) SLEDAI-2K descriptors and (B) SDI items grouped by organ system at baseline, stratified by cluster. SLEDAI-2K: SLEDAI 2000; SDI: SLICC/ACR Damage Index

Figure 4.

Trajectories of patient clusters over time. Depicted are SLEDAI-2K score trajectories (A), mean prednisone equivalent dose (B), LLDAS (C) and DORIS remission (D) from baseline until the end of the follow-up period across clusters. DORIS: definitions of remission in SLE; LLDAS: lupus low disease activity state; SLEDAI-2K: SLEDAI 2000

Figure 5.

Cox regression applied to the population stratified by clusters. Forest plots illustrating results from Cox regression models investigating clusters in relation to sustained LLDAS (A) or sustained DORIS remission (B) attainment, adjusting for age, ethnicity, sex and belimumab use, and models investigating belimumab use in relation to sustained LLDAS (C) or sustained DORIS remission attainment (D) stratified by clusters. DORIS: definitions of remission in SLE; LLDAS: lupus low disease activity state