A generator-matrix causal-inference framework separates measurable aging biomarkers from mortality-driving latent dynamics in humans
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
A central challenge in computational geroscience is to distinguish molecular quantities that predict mortality from those that causally drive it. Epigenetic clocks and aging biomarkers are increasingly used as if they were that mechanism, yet this is rarely tested directly. This distinction also bears on competing theories of aging: damage/reliability (A), hyperfunction/mTOR–IIS (B-1), and information loss (B-2). Although individual aging proteins have been tested piecemeal, no study has asked, in one framework, what fraction of mortality is measurable, whether it is causal, and whether it is reversible. Using only public, de-identified data, we evaluate this three ways. First, a Markov generator-matrix model of hallmark-load dynamics with death as an absorbing state, fitted by Bayesian inference through a joint biomarker-and-mortality likelihood to NHANES with linked mortality (n=23,844) and replicated in the Health and Retirement Study (HRS), decomposes Gompertz acceleration into visible (measured-biomarker-driven) and latent components. Second, a positive-control-calibrated, two-platform cis-pQTL Mendelian-randomization and colocalization design (UKB-PPP, de-CODE) against parental-lifespan GWAS tests whether the latent’s measurable components are causal. Third, a clock battery (Horvath, chronological; DamAge, causality-enriched damage) tests reversibility in cellular reprogramming. Within the model, ∼92% of Gompertz acceleration is assigned to a latent component not captured by measured blood-biomarker axes (NHANES 92.5%, HRS 91.6%); the latent is partly encoded in DNA-methylation signatures but not transcription. The known causal proteins are detected (LPA p=9×10⁻¹²; IL6R p=2.8×10⁻⁵), yet the latent’s components, across inflammatory, renal and growth-signalling (IGFBP3, IGF-1) axes, are null and do not colocalize on either platform. Reprogramming reverses the chronological clock (−11 to −22 yr) but not the causality-enriched damage clock. The model-inferred mortality-driving component is largely latent to accessible biomarkers; its measurable molecular proxies show no supported causal effect where the design detects known causes; and the causality-enriched damage-clock signal is resistant to partial reprogramming.
Author summary
Aging “clocks” and blood biomarkers are increasingly used as if they measured the process that makes us age and die, yet a thermometer reflects a fever without causing it. We asked whether the molecular quantities we can measure are the cause of mortality or merely downstream markers , and built a computational framework that tests this in three independent ways. First, a dynamical reliability model that treats death as an absorbing state, fitted to two large population datasets, indicated that roughly 92% of how fast mortality rises with age is not captured by the blood biomarkers we can measure. Second, a genetic causal test (a natural experiment using inherited gene variants), calibrated with proteins already known to affect lifespan, showed that the measurable markers, across inflammation, growth signalling and the epigenetic clock, do not causally change human lifespan, whereas the known-causal proteins do. Third, cellular “rejuvenation” reversed a chronological clock but not a causality-enriched damage clock. Collectively, these results indicate that popular aging measures track mortality risk rather than its cause , and we provide a reusable computational framework for telling the two apart.
