HallmarkAge: Binarized Hallmark-Aware Transcriptomic Clocks to Discover Aging Mechanisms

Aging is frequently described through mechanistic “hallmarks,” sets of pathways that differ across studies and outline separate aging mechanisms. Such classification may help interpret aging clock scores and generalize across cell-types, cohorts, and sequencing technologies. We introduce HallmarkAge, a framework that converts various hallmark gene sets into supervised binary hallmark state labels via score thresholding and trains probabilistic hallmark state classifiers from gene expression. We employ seven gene-set frameworks, producing a compact and interpretable hallmark representation that supports hallmark-space embedding, clustering, and hallmark-aware age prediction. We evaluate HallmarkAge on a single-cell and single-nucleus RNA-sequencing human skeletal muscle aging atlas across multiple cellular subsets. Hallmark-space reveals structured heterogeneity across aging states and clustering produces mechanistically readable programs with coherent marker genes. We train aging clocks using hallmark-derived features, selecting models under donor-level splits and aggregating predictions to donor-level estimates. Residual analysis highlights previously undescribed aging signals, nominating targets for perturbation screening and validation.