Adiv A. Johnson

Biological age describes how a person is aging on molecular, cellular, and physiological levels. At the present time, biological age is not something that can be fully quantified and aging itself is far too complicated to be captured by a single layer of information. As such, it is relegated to being an abstract and somewhat vague idea. In this sense, biological age is remarkably similar to health. Health, like biological age, is not well defined and too multifarious to be summarized with a single biomarker. Given their similarities, it is intriguing to consider whether or not biological age and health are synonyms. Indeed, these two ideas can be cleanly swapped with each other in a variety of scenarios. If someone appears uniquely vigorous and capable for their chronological age, that person can be described as either quite healthy or being biologically youthful. However, there are situations where biological age and health diverge from each other. It is possible, for instance, to have a short healthspan but a long lifespan. While biological age is otherwise comparable to health as an abstract idea, it is possible that biological age and health will be confidently quantifiable in the future. Indeed, such quantification may be unlocked by the integration of multimodal data and artificial intelligence. In such a scenario, biological age offers a clear advantage over health in that it operates on a more intuitive, familiar scale as opposed to an artificially created index.

1. An aging clock is any computational model that predicts age given molecular expression data [1]. Proteomic aging clocks utilize proteins as inputs [2], transcriptomic aging clocks employ gene ex...

Epigenetic aging clocks estimate age from DNA methylation patterns and have become central tools in longevity research. More recently, next-generation clocks have been developed to better compensate for the known divergence between chronological age and epigenetic age in ways that relate to lifestyle, health, and age-related disease. Although epigenetic clocks represent investigational biomarkers, these newer models are more strongly associated with all-cause mortality risk than first-generation clocks. As such, interventions that modify them are of interest. To test this, we performed a series of systematic searches and identified 41 human studies reporting the effects of interventions on at least one next-generation epigenetic clock. Our data suggest that a diverse range of pharmaceutical, lifestyle, supplementation, non-pharmaceutical clinical, and psychosocial interventions can decrease epigenetic age, including exercise, a plant-rich diet, the GLP-1 receptor agonist semaglutide, caloric restriction, ketamine, omega-3 fatty acids, a multivitamin-multimineral supplement, umbilical cord plasma, and the cholesterol-lowering drug pitavastatin. Nicotinamide riboside, rapamycin, senolytics, and several other interventions showed no detectable effect, whereas plasmapheresis and other therapeutics accelerated epigenetic aging. We also summarize reported effect sizes and compare next-generation clocks with respect to their frequency of use and responsiveness to intervention.

In human blood, it has been demonstrated that methylomic information can be used to predict smoking status, alcohol intake, and chronological age. While it is possible to robustly predict chronological age using DNA methylation information derived from buccal tissue, it remains to be determined if other variables can be directly predicted in cheek swabs. Here, we demonstrate that classifiers for smoking status and race/ethnicity can be built in a buccal methylomic dataset derived from 8,045 adults spanning an age range of 18–93 years. Furthermore, we build novel regressors for body mass index, alcohol intake, and chronological age. For each of these models, we identify the 1,000 most important CpGs and perform enrichment analyses on them to expose associated biological pathways and transcription factor targets. We additionally explore how the architecture of an epigenetic aging clock–specifically how many hidden layers are present–influences model accuracy. Finally, we build proof-of-concept, explainable deep learning models that connect DNA methylation sites annotated to genes to Reactome pathways or to transcription factors. These pathways and target sets are then used to estimate age, a feature that provides interpretability. All together, these findings further emphasize the usability of buccal data for epigenetic predictions.

A recent publication presented the results of a survey given to longevity experts that attended the 2022 Systems Aging Gordon Research Conference. Corroborating results from a prior survey and broader discussions in the field, the authors found that there was little consensus on fundamental topics, including what causes aging. While this disagreement is not particularly surprising, the spread of answers to the following prompt was unexpected: “The average lifespan in developed countries will be increased by >10 years in the next 20 years.” On a five-point agreeability scale, a slight, significant majority of respondents indicated agreement with this projection. To get a sense of how ambitious this prediction is, a previous analysis concluded that preventing every single cancer death would only increase life expectancy at birth by around three years. Moreover, the majority of gains in life expectancy since 1900 are due to improvements in preventing and managing infectious and contagious diseases. Given all of this, the lengthy lag between discovery and translation, the time required to prepare for and launch longevity-oriented clinical trials, and the limited number of reproducible interventions that robustly extend lifespan in male and female mice, it is unlikely that 10+ year increases to average lifespan are just around the corner in developed countries. While aging is undoubtedly modifiable and further improvements in longevity are achievable, the field’s optimism should not distract from what is realistically attainable and the amount of work and time required to successfully identify, test, and translate pro-longevity interventions.

Using 100 technical replicate samples from two adult buccal cohorts, we compared technical methylation variability and signal strength between the Infinium MethylationEPIC v2.0 array and the Twist Human Methylome Panel across 753,648 shared CpGs. Twist methylation sequencing showed skewed methylation distributions and fewer highly correlated CpGs than MethylationEPIC arrays. Variance analysis revealed a skew toward higher signal strength in MethylationEPIC datasets, with a subset of CpGs showing high signal strength in both methylation sequencing and array datasets. Despite these biases, four principal component (PC) trained epigenetic clocks (pcHorvath1, pcHorvath2, pcHannum, and pcDNAm PhenoAge) were robust across both technologies, even with missing data. While pcHannum and pcDNAm PhenoAge were similarly reproducible with mean absolute replicate difference (MRD) values ranging from 1.014 years to 1.194 years, pcHorvath1 was more reproducible in arrays (MRD = 0.459 years) than methylation sequencing (MRD = 2.320 years) and pcHorvath2 was more reproducible in methylation sequencing (MRD = 0.760 years) than arrays (MRD = 1.011 years). Furthermore, original non-PC versions of these clocks were less reproducible in Twist datasets and, as an example of this, MRD for uncorrected clocks went as high as 15.498 years in arrays and as high as 20.180 years in methylation sequencing. Obvious differences in age prediction were also observed in original clocks compared to their PC-trained versions across both technologies (with a mean absolute difference ranging from 4.492 years to 46.724 years). This underscores the need for careful selection of epigenetic clocks and technology-specific adjustments when optimizing for accuracy and reproducibility.

Aging biomarkers that predict age given methylomic data are referred to as epigenetic aging clocks. While the earliest, first-generation clocks were exclusively trained to predict chronological age, more recent next-generation models have been explicitly trained to associate with health, lifestyle, and/or age-related outcomes. Although these next-generation models have been trained using distinct approaches and techniques, existing evidence indicates that they associate with a greater number of health and disease signals than first-generation clocks. Moreover, they are often more predictive of age-related outcomes and appear more responsive to interventions. In this work, we provide definitions for first- versus next-generation clocks and discuss the potential merits of further dividing next-generation clocks into sub-categories. In addition, we summarize existing next-generation epigenetic aging clocks, including how they were trained and how they can be accessed. Given the relative value of interventional data over observational data, we comprehensively tabulate existing literature documenting the ability of an intervention to influence at least one epigenetic aging clock. While we acknowledge that the decision to a use a specific clock is ultimately dependent on the research application and goal, current evidence suggests that next-generation models should be generally prioritized for health-oriented association and interventional studies.