September 3, 2026 | Epigenetic methylation markers on your DNA can be used to predict your biological age, how soon you’ll die, or the development of various age-related diseases and morbidities. However, popular prediction algorithms were trained on distinct endpoints, and it is unclear fundamentally how and why each methylation profile correlates with a specific outcome. Now, researchers at the University of Southern California have extended these tests by looking at all of the gene transcription taking place as a function of each of five different popular clock predictors, increasing our understanding of the mechanisms. The work was published July in npj Aging (DOI: 10.1038/s41514-026-00446-x).
Although the gene expression changes were largely distinct among the clocks, “they do converge on four fundamental features of aging: metabolic, developmental, immune, and regulatory processes; there is some overlap in the overarching features,” explained first author Em Arpawong, Research Associate Professor of Gerontology at the USC Leonard Davis School of Gerontology.
Diagnostics World News spoke with Arpawong to better understand how transcriptomics pushes the field forward by increasing the precision and interpretability of epigenetic clocks, and to learn whether you should order a “true age” test any time soon.
Epigenetic Clock Backstory
The association between the epigenetic marker cytosine-5 methylation (occurring at CpG sites) and aging dates to the 1990s, as researchers noticed methylation increasing in some regions while decreasing in others. Epigenetic changes, including CpG methylation, influence gene expression without changing the underlying DNA sequence.
In 2013, a group published the first widely recognized epigenetic clock, looking at age as a function of CpG methylation in blood samples. Co-first author Gregory Hannum received the namesake for this predictor, which ultimately distilled age prediction to analysis of just 71 methylation marks and showed an error of 3.9 years.
Steve Horvath created a similar first-generation clock later in 2013 by looking at data from 51 different healthy tissue types. His model considered methylation for 353 CpGs, and it could predict age using almost any tissue sample with an error of 3.6 years. He also showed that cancer samples exhibit an average age acceleration of 36 years.
In effect, these first-generation clocks showed that DNA methylation is a molecular record of passing time, and in theory, any individual could measure their biological age to assess the possibility of accelerated aging.
Since Hannum and Horvath, many other epigenetic clocks have been developed, and three more popular clocks were considered in Arpawong’s study: PhenoAge, GrimAge, and DunedinPACE. PhenoAge (2018) considered phenotypic age by adding in clinical biomarkers such as glucose and creatinine levels, landing on 513 useful CpGs and improving mortality and aging-related morbidity predictions. GrimAge (2019) took a slightly different approach, first identifying DNA-methylation correlations with cumulative smoking or the levels of certain plasma proteins already associated with morbidity and mortality, and then determining how those methylation marks related to time to death. This clock included 1030 CpGs and outperformed prior versions in predicting time to death with additional measures such as time to cardiovascular disease. Interestingly, the DNA methylation surrogate for smoking outperformed the actual biomarker of smoking pack-years. Finally, DunedinPACE (2022), sometimes considered a third-generation clock, was developed to look at the rate of aging by longitudinally analyzing the same individuals over four different measurements spanning a total of 19 years. This metric includes 173 CpGs and identified a wide range of aging speed in their cohort, from 0.4 years per chronological year for one individual to 2.44 years per chronological year for another.
Add a Transcription Layer
Against this backdrop, Arpawong and her colleagues wanted to look a little deeper, not just at the methylation markers merely associated with aging or aging-related conditions but at the specific genes and biological processes involved. “We really don't know what these clocks are measuring,” Arpawong explained. “Is it normative aging? Is it pathological, or what’s really being captured? And since they've been built on machine learning algorithms, it's been relatively unclear,” she added. The clocks are sometimes thought of as “black boxes” with ambiguous explainability.
For their study, Arpawong and colleagues interrogated a dataset from the U.S. Health and Retirement Study Venous Blood Study, for which the samples included both CpG methylation and gene transcription profiles. The researchers performed one analysis for each of the five clocks and assessed which genes’ expression levels were correlated with that clock prediction. Unlike a traditional differential gene expression analysis where one predicts how expression affects a binary outcome, “we actually use the clocks as continuous variables,” Arpawong said. “One of our co-authors and collaborators, Steve Cole, started using an approach that we implemented in this study, where we used a clock as the independent predictor variable, and the gene expression as the dependent variable,” she explained. In other words, as the cumulative clock score increases, does an individual gene’s expression significantly change?
At a very basic level, increased CpG methylation occurring within gene promoters and enhancers is generally associated with a more closed chromatin state and reduced transcription from the nearby gene. However, reality is more complex. “We don’t know the true regulatory role of that DNA methylation on what it’s affecting, whether it’s something proximal or distal,” Arpawong said, and the study revealed a disconnect between affected genes and the CpG sites in each clock. For instance, as methylation changed across the 353 CpGs in the Horvath clock, only 49 genes were differentially expressed. In contrast, the 173 CpG sites from DunedinPACE yielded 3,204 differentially expressed genes (DEGs). In all cases, the vast majority (93% or more) of DEGs were not right next to a CpG site from the clock.
Additionally, the DEGs identified among the five clocks were pretty distinct. No DEGs were common among all five, with 25 common to four. The greatest overlap occurred among the 2nd and 3rd generation clocks; GrimAge and DunedinPACE shared the most, with 549 DEGs. This lack of strong overlap makes some sense, however, given the clock origins. “The clocks are capturing some of the aging processes that largely involve different genes and molecular pathways and that reflect how each clock was originally designed, so the clocks are not necessarily interchangeable or really tracking on the same processes,” Arpawong said.
Although the clocks correlated with distinct genetic expression profiles, “we can look at a more granular level at those underlying biological processes,” Arpawong said. They categorized genes according to biological process using gene ontology terms, at which point some common themes emerged. This is where the researchers identified commonality around the metabolic, developmental, immune, and regulatory processes. The manuscript explained it this way: “while broad aging-related processes (e.g., inflammation, metabolism) are recurrent, the underlying gene-level architecture and relative pathway contributions differ substantially between clocks.”
The common immune and regulatory processes include the relatively hot-topic area of general inflammation, and Arpawong had a bit more to say about how epigenetics could provide a useful metric here. For inflammation, she said, “that’s not something that’s routinely measured; that’s something that I think epigenetics [and transcriptomics] opens up the way for, if we can get even closer to the current processes that are happening with people, we can say ‘you are at the height of an inflammatory process’ … and then that’s kind of a risk indicator that we can start to implement.”
The researchers also created new clock models based on the gene transcription findings. While one might expect this approach to simply recapitulate epigenetic clock predictions, in some cases they went further. “We found that the gene expression scores that we derived from the clocks—transcriptomic aging gene scores (TAGS)—on a moderate level [they] track with their parent clocks, and they predict some outcomes that are often used to indicate accelerated aging, like frailty, disability, cardiovascular health, diabetes status, inflammation, and mortality, the same or more strongly than the DNA methylation scores alone,” Arpawong said. “The idea,” she said, “is that at the transcript level we’re getting something that is more immediate to processes occurring in an individual, and that may be a little bit closer to their current health state or profile.” In contrast, the epigenetic marks are more stable changes that have happened over a lifetime.
“These TAGS essentially create a new tool for the field where the epigenetics is just not currently available,” Arpawong said, pointing to various cancer studies as an example.
The Future of Epigenetic and Multimodal Clocks
For now, these clocks are mostly a research tool, not quite ready for personalized medicine. Your doctor is unlikely to order one, but some companies do offer direct testing, with out-of-pocket pricing on the order of $300-500. More often, “at a group and a population level, these clocks have been used to predict outcomes like mortality,” Arpawong said. “There is some implementation and testing in intervention studies, but I think there’s that difficulty of having a margin of error of four or five years,” she added, which makes it difficult to reliably determine whether an intervention is affecting the clock profile.
Arpawong says a few things need to happen in the field. One is to “increase the precision in these clocks and how to construct them, maybe not with just epigenetics, but with other layers of data.” For instance, “some people are working on things like metabolomics, proteomics, building in a lot of different layers including microbiome or integrating measured biomarkers that you would typically do in a doctor’s office, for example … seeing whether or not we can use multiple data types to really get at a better individual level predictor.”
A second major push is to increase interpretability. Arpawong was first author on another paper earlier this year reporting PhysAge, a clock incorporating additional biomarkers. “The idea behind that clock is that it is an epigenetic clock and score overall that will give you a physiological health age in the metric of years,” she said. “But once you get that number, you can also look at the component parts that went into the score.” For instance, “one of the components is cholesterol, another one is peak flow for lung function, another one is CRP for inflammation,” she said. “That lends towards more of the interpretability, which would hopefully be more usable in a clinical setting.”
Arpawong imagines a future where the clock predictions become more standard. “In pediatrics, we have these growth charts for kids; every clinician and parent knows what that is,” she said. “We could do something similar with these clocks … a biological clock for adults … we can establish this developmental chart or age chart to see where people fall with their biological aging, and we can set these norms from what we’re seeing across populations.” Ideally, one could then look at the specific systems influencing the measure and “create these more trackable and interpretable processes that get implemented in a clinic and can give feedback to a patient immediately,” she said.
She explained how epigenetics can provide a unique window into biological state and risk, especially coupled with the enhanced understanding provided by transcriptional data. “One of the advantages of the epigenetics is that … they’re encapsulating risk profile, much more so than one biomarker could really do.” For instance, “if somebody has a normal HbA1c and they don’t have diabetes, it doesn’t necessarily mean they’re not at risk for it, and based on other lifestyle or exposures or other factors, the epigenetics seemingly is giving us more of a risk predictor versus the standard biomarkers.”