A deep-learning based biomarker of systemic cellular senescence burden to predict mortality and health outcomes.

Introduction: The accumulation of senescent cells is a recognized hallmark of biological aging and is associated with the onset of multiple chronic medical conditions. Senescent cells exhibit a distinct secretory profile, known as the senescence-associated secretory phenotype (SASP), which can propagate cellular senescence to neighboring and distant tissues. Measuring SASP factors in blood serves as a practical proxy for cellular senescence burden and may help track disease states and intervention outcomes. Methods: We developed and validated a composite SASP Score by integrating large-scale population proteomics data with a semi-supervised deep learning framework. The analytical workflow included: (1) selection of biologically curated SASP proteins; (2) development of a Guided autoencoder with Transformer (GAET) model using data from the UK Biobank Pharma Proteomics Project (UKB-PPP); (3) internal evaluation and association analyses within the UK Biobank; and (4) external validation and longitudinal assessment in an independent randomized clinical trial cohort. Results: The deep learning-based SASP Score was a strong, independent predictor of mortality risk and incident serious, chronic medical conditions (e.g., dementia, COPD, myocardial infarction, stroke). In an independent cohort, multimodal exercise significantly changed the SASP Score trajectory over 18 months. Discussion: Our findings support the potential of a deep learning-derived SASP Score as a biomarker for systemic cellular senescence burden. Our statistical approach can offer enhanced interpretability and cross-platform utility, providing a valuable tool for aging research and the evaluation of geroscience-guided interventions.