Insights to aging prediction with AI based epigenetic clocks

Abstract

Over the past century, human lifespan has increased remarkably, yet the inevitability of aging persists. The disparity between biological age, which reflects pathological deterioration and disease, and chronological age, indicative of normal aging, has driven prior research focused on identifying mechanisms that could inform interventions to reverse excessive age-related deterioration and reduce morbidity and mortality. DNA methylation has emerged as an important predictor of age, leading to the development of epigenetic clocks that quantify the extent of pathological deterioration beyond what is typically expected for a given age. Machine learning technologies offer promising avenues to enhance our understanding of the biological mechanisms governing aging by further elucidating the gap between biological and chronological ages. This perspective article examines current algorithmic approaches to epigenetic clocks, explores the use of machine learning for age estimation from DNA methylation, and discusses how refining the interpretation of ML methods and tailoring their inferences for specific patient populations and cell types can amplify the utility of these technologies in age prediction. By harnessing insights from machine learning, we are well-positioned to effectively adapt, customize and personalize interventions aimed at aging.

Keywords: DNA methylation; aging; artificial intelligence; classification and regression trees; clocks; deep learning; epigenetics; explainability.

Figure 1.

Epigenetic clocks and future advances. (a) Biological vs. Chronological age to illustrate the concept of age acceleration: on the left, a datapoint has been added for a specific patient representing where biological age surpasses chronological age suggests aging processes that is faster than expected. Conversely, in the middle, a biological age less than chronological age for a specific patient indicates the influence of biological processes and lifestyle factors that contribute to a slower aging process. On the right, the goal of many interventions is to slow down the process of aging for a specific patient as they age, represented in this graphic by the patient (data point) moving from an accelerated aging state to a decelerated over time. (b) Age prediction methods: graphical representation of classical methods such as EWAS (1 CpG to predict age as indicated using green arrow), penalized multivariable regression (e.g., LASSO, elastic Net; penalized approach models many CpGs simultaneously while selecting a subset that are the most predictive of age), and advanced machine learning techniques like random forest and deep learning (graphic at bottom left). These advanced methods can further minimize the residual between biological and chronological ages with the potential to provide insights into the underlying mechanisms. This is represented graphically by plotting the same patient twice as a two data points – the traditional regression approaches (blue arrow/line) estimate a larger residual between biological and chronological age, whereas the machine learning approaches estimate a smaller residual (red arrow/line). (c) Explainable AI techniques: depicts the use of explainable AI to organize CpGs into genes and pathways within deep learning models, enhancing the interpretability of epigenetic data. DNAm beta values representing the proportion of methylated alleles at individual CpG sites are aggregated and transformed based on related genes to form gene summaries of DNAm. Similarly, gene-level DNAm information is summed and aggregated on the pathway level. Green arrows indicate this mapping between CpGs to genes to pathways. (d) CpG and Cell Type Interactions: discusses the integration of CpG sites with cell type data estimated through deconvolution approaches. Beta values at specific CpG sites for one patient are represented using the gray squares in the upper left of the panel. Cell type abundance for the same patient is represented through the circles in the upper right of the panel, with different colors for each cell type. This panel explores how CART models can elucidate molecular mechanisms behind age acceleration specific to certain cells or patient subpopulations when derived interactions are applied to generalized linear modeling approaches. The decision tree selects specific cell types and CpGs from which decision rules are made, further binning patients based on these decisions. The green arrows indicate the logical flow between the splits. If a decision split based on a cell-type follows a CpG decision split and vice versa, there exists a potential statistical interaction between a CpG and cell-type. These interactions are tested through statistical modeling on the right, which demonstrates how the association between a specific CpG’s methylation status and age can change depending on the cell-type (i.e., different slope for blue versus yellow cell-type).