Tu114 - Decoding the Complexity of Childhood-onset Systemic Lupus Erythematosus (cSLE) Using a Multi-dimensional Approach: Enhancing Theragnosis with an Immunological-based Diagnostic Model Revealing Mechanistic Insights

Joo Guan Yeo – Translational Immunology Institute, Singhealth Duke-NUS Academic Medical Centre; Martin Wasser – Translational Immunology Institute, Singhealth Duke-NUS Academic Medical Centre; Su Li Poh – Translational Immunology Institute, Singhealth Duke-NUS Academic Medical Centre; Kai Liang Teh – KK Women’s and Children’s Hospital; Lena Das – KK Women’s and Children’s Hospital; Yun Xin Book – KK Women’s and Children’s Hospital; Nicholas Kim Huat Khoo – Translational Immunology Institute, Singhealth Duke-NUS Academic Medical Centre; Zoe Mei Yun Tay – Translational Immunology Institute, Singhealth Duke-NUS Academic Medical Centre; Rachel Cheong – Translational Immunology Institute, Singhealth Duke-NUS Academic Medical Centre; Judith Ju Ming Wong – KK Women’s and Children’s Hospital; Thaschawee Arkachaisri – KK Women’s & Children’s Hospital; Salvatore Albani – Translational Immunology Institute, Singhealth Duke-NUS Academic Medical Centre

Abstract Text: Childhood-onset systemic lupus erythematosus(cSLE) is usually more severe than adult-onset SLE. We aim to resolve the immunopathogenic complexity of SLE using high-dimensional mass cytometry(CyTOF). Machine-learning and mechanistically relevant cell clusters were used to create a diagnostic model.

For discovery, peripheral blood mononuclear cells(PBMCs) from 30 cSLE patients (53 timepoints, median age:14 years) and 17 age-matched healthy controls were studied. Unsupervised analysis was done using our Extended Poly-dimensional Immunome Characterisation(EPIC) pipeline. Discovery cohort data identified predictive cell clusters using machine-learning(linear SVM, PLS-DA, random forest (RF)) with MetaboAnalyst v6.0. Cell frequencies are expressed as CD45+ PBMCs percentage (median[interquartile range]) with statistical significance p< 0.05(Mann-Whitney U). For independent validation, 18 cSLE (median age:11.5 years) and 23 healthy individuals were used.

67 unique cell clusters formed the basis to map validation cohort data. Mapping was automated using previous unsupervised analysis and expert annotation of discovery data. A memory Treg-like population (CD3+CD4+CD45RO+CD25-Foxp3+CTLA+) was higher in cSLE than healthy (2.75[1.90-4.36] vs. 1.28[0.83-1.78],p < 0.0001). Concurrently, a CD8+CD45RA+BAFF+ T-cell population was enriched in cSLE (8.91[6.78-11.4] vs. 3.22[2.79-4.86],p < 0.0001).

These populations and eight others were selected through collective ranking of the most important features generated by machine-learning algorithms on discovery cohort data to create a classification model. The average accuracy of discriminating cSLE and healthy based on 100 cross validations in the discovery cohort is 85.6%[85.4-85.7], with 88.9%[83.3-100] sensitivity for the validation cohort. Further studies will quantitate autoantibody titers with ELISA for comparison.

In conclusion, this immunological-based classifier model can augment current clinical criteria while imparting a theragnostic perspective to cSLE.