A deep neural network provides an ultraprecise transcriptomic clock of the short-lived fish Nothobranchius furzeri and identifies predictive genes translatable to human aging.

Abstract

A key and unresolved question in aging research is how to quantify aging at the individual level that led to development of "aging clocks", machine learning algorhythms trained to predict individual age from high-dimensional molecular data under the the assumption that individual deviations of the predicted age from the chronological age contain information on the individual condition (often referred to as "biological age"). A full validation of such clocks as biomarkers for clinical studies of ageing would require a comparison of their predictions with information on actual lifespan and long-term health. Such studies take decades in humans, but could be conducted in a much shorter time-frame in animal models. We developed a transcriptomic clock in the turquoise killifish Nothobranchius furzeri. This species is the shortest-lived vertebrate that can be cultured in captivity and is an emerging model organism for genetic and experimental studies on aging. We developed a proprietary deep learning architecture that autonomously selects a customizable number of input genes to use for its predictions in order to reduce overfitting and increase interpretability, and adopts an adversarial learning framework to identify tissue-independent transcriptional patterns. We called this architecture the Selective Adversarial Deep Neural Network (SA-DNN) and trained it on a multi-tissue transcriptomic dataset of N. furzeri. This SA-DNN predicted age of the test set with an accuracy of 1 day, i.e. less than 1% of the total species lifespan and detected genetic, pharmacological and environmental interventions that are known to influence lifespan in this species. Finally, a human transcriptomic multi-tissue clock that uses as input the orthologs of the genes selected by our SA-DNN in N. furzeri reaches an average error of[~] 3 years rivalling epigenetic clocks. Our SA-DNN represents the prototype of a new class of aging clocks that provide biomarkers applicable to intervention studies in model organisms and humans.