Understanding Aging at Multi-scale Using Explainable AI

Abstract

As human lifespans increase, understanding the biological and clinical mechanisms that shape aging has become increasingly important. This dissertation presents a set of explainable AI (XAI) frameworks that illuminate aging at multiple scales, ranging from population-level health data to bulk transcriptomics and single-cell gene expression. I begin with IMPACT, an XAI framework for all-cause mortality prediction in NHANES dataset. IMPACT improves prediction accuracy over traditional models and uses XAI methods to reveal previously underappreciated risk factors and clinically meaningful feature interactions. Building on this foundation, ENABL Age extends the IMPACT framework to model biological age. ENABL Age combines machine learning with XAI to estimate biological age and to quantify how specific lifestyle, clinical, and biochemical factors contribute to accelerated or slowed aging. This framework provides individualized insights into modifiable components of aging and supports the development of interpretable precision aging tools. At the molecular scale, DeepProfile learns biologically meaningful latent representations from 50,211 cancer transcriptomes across 18 tumor types. It identifies universal immune activation signals, cancer-type specific subtype structure, and mechanistic links among mutation burden, cell-cycle activity, antigen presentation, and patient survival. By studying cancer across many organs, DeepProfile also offers insight into organ health and organ aging, illustrating how unsupervised learning can uncover clinically relevant biology from large transcriptomic datasets. Finally, ACE is an explainable deep generative model for single-cell RNA sequencing data that isolates aging-related gene expression changes from dominant background variation, enabling the study of cellular aging. Applied to mouse, fly, and human datasets, ACE recovers tissue and cell-type specific aging signatures, identifies conserved aging pathways across species, predicts biological age at cellular resolution, and prioritizes novel regulators such as Uba52, whose relevance is validated through lifespan-shortening RNAi experiments in C. elegans. Together, these contributions form an integrated XAI-driven framework for understanding aging at multi-scale and advance both mechanistic aging biology and transparent approaches for improving human healthspan.