Single-cell transcriptomic integrated with machine learning reveals retinal cell-specific biomarkers in diabetic retinopathy

Diabetic retinopathy (DR) remains a principal cause of vision impairment worldwide, involved complex retinal cellular pathophysiology that remains incompletely understood. To elucidate cell-type-specific molecular signatures underlying DR, we generated a high-resolution single-cell transcriptomic atlas of 297,121 retinal cells from 20 Chinese donors, including non-diabetic controls (26.4%), diabetic without retinopathy (23.4%) and DR (50.2%). Following rigorous quality control, batch-effect correction, and clustering and annotation, 10 major retinal cell populations were delineated. Differential expression analyses across disease states within each cell type yielded candidate gene sets, which were further refined via a multi-stage machine-learning pipeline combining L1-regularized logistic regression and recursive feature elimination with cross-validation, alongside bootstrap stability selection. Resulting cell-type-specific classifiers achieved high accuracy (79-95%) and AUCs (0.85-0.99) in distinguishing DR disease states. Enrichment analyses implicated immune activation, oxidative stress, neurodegeneration and synaptic dysfunction pathways across multiple cell types in retina. Integrating 567 unique marker genes from all cell types, a general multilayer perceptron classifier achieved 95.31% overall accuracy on held-out test data, demonstrating the translational potential of these signatures for non-diabetic controls, diabetic without retinopathy and DR classification. This high-resolution atlas and the accompanying analytic framework provide a robust computational framework for biomarker discovery, mechanistic insight and targeted intervention strategies in diabetic retinal diseases.