Improved methods for age, frailty, and mortality prediction in large-scale longevity studies

Abstract

Aging is an extremely complex process that lacks a consensus definition, making it difficult to study. The most definitive way to study aging is to assess differences in group lifespan, which for short-lived organisms is routine but for mammals is highly inefficient. To solve this problem, researchers have developed machine-learning-based biomarkers for aging—often referred to as “clocks”—which typically use high-throughput biomolecule assays to predict age or mortality. However, these aging clocks have failed to achieve their full potential because they lack scalability, are costly, and have not been developed to the same extent in mice, the primary research mammal. In Chapter 1, I examine the difficulty of defining aging, present the theory that heritable information loss drives aging, and introduce the promise and limitations of aging biomarkers. In Chapter 2, I describe the development of TIME-Seq, a novel method for highly-efficient DNA methylation sequencing that enables large-scale epigenetic age experiments. My collaborators and I used TIME-Seq to train and validate many epigenetic clocks in multiple tissue and cell types from thousands of humans and mice. We measured the association between our clocks and diverse phenotypes, and we used TIME-Seq to assess interventions known to slow and accelerate aspects of aging. Leveraging the scalability of our approach, we tracked epigenetic age deceleration with high temporal resolution in cohorts of mice treated with lifespan-extending drugs. In parallel, we discovered an even more efficient approach for age prediction from shallow sequencing of TIME-Seq libraries. In Chapter 3, I describe the development of novel phenotypic age and mortality clocks for mice. My collaborators and I tracked a large cohort of mice until death, collecting phenotypic data and biosamples from them at up to five timepoints. We applied this data to develop the first mouse phenotypic age model (mPhenoAge). With blood-derived DNA methylation data and plasma metabolomic data, we trained highly accurate predictors of mPhenoAge as well as mortality, frailty, and chronological age. In summary, my research led to new, more efficient, and more useful methods to predict age, frailty, mortality, and phenotypic age. With these methods, we probed questions of basic aging biology and present novel experimental strategies that promise to accelerate longevity intervention testing.