Unsupervised learning of aging principles from longitudinal data

Aging is a major risk factor for numerous diseases and mortality, yet the relationship between age-related physiological changes and lifespan remains poorly understood. Current biomarkers often rely on supervised models requiring chronological age or mortality data for training, limiting their applicability. Alternatively, frailty indices, reflecting accumulated health deficits, correlate with lifespan but lack a mechanistic understanding. This research proposes that aging is a dynamic system nearing instability, driven by a few key features. By leveraging the Mouse Phenome Database (MPD), a rich source of longitudinal data, the researchers developed a deep learning approach to identify quantitative aging principles and their connection to mortality. The core hypothesis posits that aging's manifestation across multiple levels of organismal organization can be captured by a single, slowly evolving variable that reflects the increasing instability of the system. This concept draws from dynamic systems theory and critical phenomena, where fluctuations around a critical point, indicating a system's transition from stability to instability, are amplified, resulting in exponential growth. This perspective offers a novel way to quantify aging and its relationship with mortality, moving beyond simple correlation-based approaches. The study utilized a deep neural network to identify such a variable, offering a data-driven approach to understanding aging dynamics.

Existing studies have utilized various approaches to quantify aging. Supervised machine learning models have produced biomarkers of aging but require chronological age or mortality as training labels. Alternatively, frailty indices, which aggregate health deficits, show correlations with lifespan but lack a clear mechanistic basis. This study builds upon dynamic systems theory and critical phenomena, where the transition from stability to instability in a system is characterized by fluctuations along a dominant principal component. This theoretical framework offers a novel lens through which to examine aging as a dynamic process.

The researchers used complete blood count (CBC) data from the MPD, focusing on a subset of 12 measurements from 6693 animals. Principal component analysis (PCA) was initially employed to reduce the dimensionality of the data, revealing that the first principal component (PC) score correlated strongly with chronological age and exhibited increasing variance with age, indicating a stochastic process. This finding aligned with the theoretical framework of critical dynamics, suggesting that the first PC approximates an "order parameter" reflecting the system's instability. To move beyond the limitations of linear PCA, the researchers developed a deep artificial neural network combining a denoising autoencoder (AE) and an autoregressive (AR) model (AE-AR). The AE reduced the dimensionality of the CBC data, capturing non-linear relationships among variables, while the AR model captured the temporal dynamics of the reduced representation. The AR model's output, termed the dynamic frailty indicator (dFI), serves as an approximation of the order parameter. The model was trained using a combination of longitudinal and cross-sectional data, leveraging the larger cross-sectional data to enhance the training of the AE portion of the network. The model's performance was evaluated using various metrics, including auto-correlation of dFI, its correlation with other aging biomarkers, and its predictive power for remaining lifespan. The effects of lifespan-modulating interventions (high-fat diet and rapamycin treatment) were also analyzed using the dFI.

The dFI, generated by the AE-AR model, exhibited several desirable properties of a biological age marker: It increased exponentially with age in the validation datasets, accurately predicted remaining lifespan, demonstrating strong correlations with various hallmarks of aging including the physiological frailty index (PFI), markers of inflammation (CRP and KC), and a proxy for senescent cell accumulation (p16-luciferase flux). Importantly, the dFI's increase in rate matched the mortality acceleration rate. The dFI also responded appropriately to interventions known to affect lifespan: it increased significantly in males subjected to a high-fat diet (shortening lifespan) while showing a decrease in response to rapamycin (extending lifespan). Furthermore, the study observed a late-life mortality deceleration, consistent with the model's prediction of a limiting mortality value. This was observed in both the male and female mice cohorts, providing evidence that supports the validity of the aging at criticality model. These results were also benchmarked against a supervised Cox proportional hazard model; indicating a strong concordance between results, which confirms the reliability and effectiveness of the unsupervised dFI model. Lastly, the study found that the myeloid lineage provided the most reliable predictors of biological age.

The findings provide strong support for the "aging at criticality" hypothesis, suggesting that aging is driven by the dynamic instability of regulatory networks, reflected by the exponential amplification of small fluctuations in physiological variables. The dFI, generated through the unsupervised deep learning approach, successfully identifies a single variable capturing the essence of this instability, offering a mechanistic link between physiological changes and mortality. The strong correlations between the dFI and various hallmarks of aging validate its utility as a comprehensive biomarker, transcending the limitations of simple composite measures like the frailty index. The study's success in identifying aging-related biomarkers without relying on age or mortality labels highlights the potential of unsupervised learning for discovering novel aging markers. The results suggest that interventions modulating the dFI may simultaneously alleviate various aging hallmarks.

This study introduces a novel unsupervised machine learning approach to identify aging biomarkers from longitudinal data. The resulting dynamic frailty indicator (dFI) provides a powerful, data-driven tool for understanding the aging process in mice. The dFI exhibits strong correlations with known hallmarks of aging, accurately predicts lifespan, and responds predictably to life-extending and life-shortening interventions. Future research should focus on validating the dFI in other species, including humans, and exploring its potential applications in developing and evaluating anti-aging therapies.

The study's findings are primarily based on data from mice, and the generalizability to other species, especially humans, needs further investigation. The complexity of the AE-AR model might limit its interpretability, and future work should focus on improving model transparency and uncovering the underlying biological mechanisms driving dFI changes. The study relies on complete blood count data, which might not capture the full complexity of aging processes and could benefit from the inclusion of data from other biological domains.