Deep learning-based integrative single-cell transcriptomic analysis of vitiligo

Introduction Vitiligo is an immune-mediated depigmentation disorder in which interferon-γ signalling and autoreactive CD8+ T cells are central, but the cross-cohort cellular ecosystem and donor-level disease signature remain incompletely resolved. Methods We conducted a retrospective analysis of three public skin single-cell RNA-sequencing datasets comprising 35 donors and 93,287 post-quality-control cells. We combined donor-aware cell-composition and pseudobulk analyses with a deep attention multiple-instance learning model, and reserved a separate balanced patient -control cohort for independent evaluation. Results Eleven major cell populations were identified. The donor-level classifier achieved 0.914 accuracy and an area under the receiver operating characteristic curve of 0.990 in donor-stratified five-fold cross-validation. After Benjamini -Hochberg correction across 11 cell types, vitiligo lesions showed macrophage expansion and depletion of fibroblasts, endothelial cells, and mast cells (all q = 0.029). Integrated ranking prioritised an interferon-γ/CXCL9 -CXCL10 programme, while CD8+ T cells displayed Trm-associated and cytotoxic-effector transcriptional states. An eight-feature signature trained only in the primary discovery cohort achieved an area under the curve of 0.944 (95% CI, 0.778 -1.000) and accuracy of 0.750 (95% CI, 0.468 -0.911) in a small independent cohort of six healthy controls and six patients with vitiligo; exact label permutation supported discrimination ( p = 0.003). A directional omnibus test across ten prespecified validation measures supported panel-level concordance ( p = 0.023), although no individual measure remained significant after multiple-testing correction. Discussion These findings provide preliminary external support for a multivariable signal connecting an expanded macrophage compartment, interferon-associated chemokine signalling, and Trm-like CD8+ T-cell states.