BackgroundWe previously developed a novel machine learning (ML) pipeline leveraging analysis of gene expression data to identify subsets of SLE patients with common molecular patterns of disease or endotypes.1 These molecular subsets exhibited significant differences in clinical characteristics, frequency of subsequent flares, and clinical responsiveness to a lupus biologic, tabalumab. The current study makes use of the ML classifier to determine endotype membership in an independent validation cohort of SLE patients.MethodsGene expression by RNA-sequencing of whole blood and clinical metadata were collected from 91 SLE patients from two clinical trials (NCT03626311 and NCT03180021). Patients met ACR classification criteria of SLE and patients from one trial had renal biopsies at the time gene expression was measured. Endotype membership of the 91 patients was identified using a random forest classifier trained on 2183 lupus patient transcriptomes, employing 26 modules of genes reflecting immune cell types and inflammatory processes. Lupus Cell and Immune Score (LuCIS), a continuous score measuring the extent of immune perturbations determined by ridge-penalized logistic regression, was also calculated for each patient.ResultsThe ML prediction of independent SLE patients into endotypes yielded eight subsets with molecular patterns mirroring those found previously in a development and testing cohort of 3166 patients (figure 1). Endotypes were designated A-H, with A representing the group with the least number of transcriptional lupus-related aberrancies and H representing the group with the greatest immunologic perturbations. Groups H, A, C, and E contained the greatest number of patients whereas B and G were small and underrepresented in this cohort. Endotype H was comprised of the greatest number of patients with proliferative lupus nephritis (LN) whereas no patient with LN was found in subset A or B. Serum complement differed among the subsets, with lower levels reflected in the more immunologically active subsets. LuCIS values reflected the immunologic activity of the subsets, but did not correlate with SLEDAI, although they were moderately, inversely correlated with serum C3 and C4 levels (figure 2). Eight patients had moderate/severe flares during the six months of the trials, all of whom had elevated LuCIS scores at baseline (figure 3).ConclusionA novel endotyping pipeline based on gene expression profiles and ML identified previously observed patient endotypes in new datasets. Patients in the endotypes with the least immunologic activity did not have proliferative nephritis and also experienced no lupus flares during the subsequent six months. Endotyping SLE patients based on transcriptional profiles can provide important prognostic information and provide novel molecular insights in support of personalized management.ReferenceKim YH, Park MR, Kim SY, Kim MY, Kim KW, Sohn MH. Respiratory microbiome profiles are associated with distinct inflammatory phenotype and lung function in children with asthma. J Investig Allergol Clin Immunol. 2023 Jun 1:0. doi: 10.18176/jiaci.0918. Epub ahead of print. PMID: 37260034.Abstract 1104 Figure 1Identification of Endotypes Among 91 SLE Patients. Molecular subsets identified by a random forest algorithm using gene set variation analysis (GSVA) enrichment scores of 26 immune/inflammatory modules. Clinical metadata for each patient (x-axis) was annotated as shown. Heatmap constructed in R using the ComplexHeatmap package.Abstract 1104 Figure 2LuCIS Correlations with Clinical Data. Pearson correlations of LuCIS values with baseline clinical characteristics. Each data point is colored by endotype membership. Plots were constructed in R using the ggplot2 package.Abstract 1104 Figure 3LuCIS of Eight Patients with Flares. Visualization of the LuCIS values at baseline in eight patients from NCT03626311 who experienced a flare over the six months of the trial. The solid black line refers to the mean +1 standard deviation of the LuCIS values of all 91 patients. Plot constructed in R using the ggplot2 package and edited in Microsoft PowerPoint.Lay SummaryLupus patients present with arrays of symptoms that are highly variable, a phenomenon called heterogeneity. Heterogeneity is also observed in the biological mechanisms that underlie lupus disease activity. To address this issue, we identified endotypes, or subsets of patients with commonalities in these underlying mechanisms. We previously developed computational algorithms to accurately predict endotype membership of any given lupus patient. Here, we are validating one such algorithm in a new, independent set of patients. We were able to identify the same subsets (endotypes) in the new data sets with differences in clinical characteristics similar to what was previously observed in the development and testing data sets. These results validate the use of gene expression profiles to provide information that could support lupus patients clinically.