Machine learning–driven immunophenotypic stratification of mixed connective tissue disease, corroborating the clinical heterogeneity

Abstract

Objective: The objective of this study was to stratify patients with MCTD, based on their immunophenotype. Methods: We analysed the immunophenotype and transcriptome of 24 immune cell subsets [from patients with MCTD, SLE, idiopathic inflammatory myopathy (IIM) and SSc] from our functional genome database, ImmuNexUT (https://www.immunexut.org/). MCTD patients were stratified by employing machine-learning models, including Random Forest, trained by immunophenotyping data from SLE, IIM and SSc patients. The transcriptomes were analysed with gene set variation analysis (GSVA), and the clinical features of the MCTD subgroups were compared. Results: This study included 215 patients, including 22 patients with MCTD. Machine-learning models, constructed to classify SLE, IIM and SSc patients, based on immunophenotyping, were applied to MCTD patients, resulting in 16 patients being classified as having an SLEimmunophenotype and 6 as having a non-SLE immunophenotype. Among the MCTD patients, patients with the SLE immunophenotype had higher proportions of Th1 cells f2.85% [interquartile range (IQR) 1.54–3.91] vs 1.33% (IQR 0.99–1.74) P ¼ 0.027g and plasmablasts [6.35% (IQR 4.17–17.49) vs 2.00% (IQR 1.20–2.80) P ¼ 0.010]. Notably, the number of SLE-related symptoms was higher in patients with the SLE immunophenotype [2.0 (IQR 1.0–2.0) vs 1.0 (IQR 1.0–1.0) P ¼ 0.038]. Moreover, the GSVA scores of interferon-α and -γ responses were significantly higher in patients with the SLE immunophenotype in central memory CD8þ T cells, while hedgehog signalling was higher in patients with the non-SLE immunophenotype, in five-cell subsets. Conclusion: This study describes the stratification of MCTD patients, based on immunophenotyping, suggesting the presence of distinct immunological processes behind the clinical subtypes of MCTD.