Identification of a blood test-based biomarker of aging through deep learning of aging trajectories in large phenotypic datasets of mice

Abstract

We proposed and characterized a novel biomarker of aging and frailty in mice trained from the large set of the most conventional, easily measured blood parameters such as Complete Blood Counts (CBC) from the open-access Mouse Phenome Database (MPD). Instead of postulating the existence of an aging clock associated with any particular subsystem of an aging organism, we assumed that aging arises cooperatively from positive feedback loops spanning across physiological compartments and leading to an organism-level instability of the underlying regulatory network. To analyze the data, we employed a deep artificial neural network including auto-encoder (AE) and auto-regression (AR) components. The AE was used for dimensionality reduction and denoising the data. The AR was used to describe the dynamics of an individual mouse's health state by means of stochastic evolution of a single organism state variable, the "dynamic frailty index" (dFI), that is the linear combination of the latent AE features and has the meaning of the total number of regulatory abnormalities developed up to the point of the measurement or, more formally, the order parameter associated with the instability. We used neither the chronological age nor the remaining lifespan of the animals while training the model. Nevertheless, dFI fully described aging on the organism level, that is it increased exponentially with age and predicted remaining lifespan. Notably, dFI correlated strongly with multiple hallmarks of aging such as physiological frailty index, indications of physical decline, molecular markers of inflammation and accumulation of senescent cells. The dynamic nature of dFI was demonstrated in mice subjected to aging acceleration by placement on a high-fat diet and aging deceleration by treatment with rapamycin.

Longitudinal analysis of blood markers reveals progressive loss of resilience and predicts human lifespan limit

“…In frail individuals, however, the intervention could produce lasting effects and reduce frailty, thus increasing lifespan beyond healthspan. This argument may be supported by longitudinal studies in mice suggesting that the organism state is dynamically unstable, the organism state fluctuations get amplified exponentially at a rate compatible with the mortality rate doubling time, and the effects of transient treatments with life-extending drugs such as rapamycin produce a lasting attenuation of frailty index 44 .…”

“…A proper identification of such a feature requires massive high-quality longitudinal measurements and sophisticated approaches auto-regressive models. In a similar study involving CBC variables of aging mice, we were able to obtain an accurate predictor associated with the age, risks of death (and the remaining lifespan), and frailty 44 . In this work we turned the reasoning around and choose to quantify the organism state by the log-linear proportional hazards estimate of the mortality rate followed 15,29,45 , using CBC and physical activity variables.…”

Longitudinal analysis of blood markers reveals progressive loss of resilience and predicts ultimate human lifespan limit

“…In frail individuals, however, the intervention could produce lasting effects and reduce frailty, thus increasing lifespan beyond healthspan. This argument may be supported by longitudinal studies in mice suggesting that the organism state is dynamically unstable, the organism state fluctuations get amplified exponentially at a rate compatible with the mortality rate doubling time, and the effects of transient treatments with life-extending drugs such as rapamycin produce a lasting attenuation of frailty index [39].…”

“…A proper identification of such a feature requires massive high quality longitudinal measurements and sophisticated approaches auto-regressive models. In a similar study involving CBC variables of aging mice, we were able to obtain an accurate predictor associated with the age, risks of death (and the remaining lifespan), and frailty [39]. In this work we turned the reasoning around and choose to quantify the organism state by the loglinear proportional hazards estimate of the mortality rate followed [15,25,40], using CBC and physical activity variables.…”

Interpretable machine learning for high-dimensional trajectories of aging health

“…This complexity is compounded by the heterogeneity and stochasticity of individual aging outcomes [3,4]. Strategies to simplify the complexity of aging include identifying key biomarkers that quantitatively assess the aging process [5,6] or integrating many variables into simple and interpretable onedimensional summary measures of the progression of aging, as with "Biological Age" [7][8][9], clinical measures such as frailty [10,11], or recent machine learning models of aging [12,13]. Nevertheless, one-dimensional measures only summarize the progression of aging, and so can miss significant aspects of high-dimensional aging trajectories and of heterogeneous aging outcomes.…”