Ensemble-DeepSets: an interpretable deep learning framework for single-cell resolution profiling of immunological aging

Abstract

Immunological aging (immunosenescence) drives increased susceptibility to infections and reduced vaccine efficacy in elderly populations. Current bulk transcriptomic aging clocks mask critical cellular heterogeneity, limiting the mechanistic dissection of immunological aging. Here, we present Ensemble-DeepSets, an interpretable deep learning framework that operates directly on single-cell transcriptomic data from peripheral blood mononuclear cells (PBMCs) to predict immunological age at the donor level. Benchmarking against 27 diverse senescence scoring metrics and existing transcriptomic clocks across four independent healthy cohorts demonstrates superior accuracy and robustness, particularly in out-of-training-distribution age groups. The model\'s multi-scale interpretability uncovers both conserved and cohort-specific aging-related gene signatures. Crucially, we reveal divergent contributions of T cell subsets (pro-youth) versus B cells and myeloid compartments (pro-aging), and utilize single-cell resolution to highlight heterogeneous aging-associated transcriptional states within these functionally distinct subsets. Application to Systemic Lupus Erythematosus (SLE) reveals accelerated immune aging linked to myeloid activation and altered myeloid subset compositions, illustrating clinical relevance. This framework provides a versatile tool for precise quantification and mechanistic dissection of immunosenescence, providing insights critical for biomarker discovery and therapeutic targeting in aging and immune-mediated diseases.