Data curation and analysis pipeline
We constructed a robust and standardized analysis pipeline for curating and analyzing a large database of interventions (Fig. 2). First, we created a list of publicly (Gene Expression Omnibus (GEO) and the European Molecular Biology Laboratory (EMBL)) and privately (TruDiagnostic) available clinical trials and studies for interventions hypothesized to promote longevity. For each study, we curated study-level metadata such as intervention type, study population, diseases in population (if any), age range and percent female. Second, we harmonized metadata in each study to have standardized columns such as: age, sex, sample ID, individual ID, follow-up time from baseline draw and sample type (control versus participant). Third, we calculated more than 110 DNAm biomarkers for each of these datasets. Fourth, we adjusted for chronological age to obtain age residuals for all DNAm biomarkers in every dataset. Fifth, we normalized the biomarker age residuals in each dataset to make them comparable across interventional studies and across biomarkers. Sixth, we performed paired t-test analysis to compare pre- and postintervention samples. Finally, we generated heat maps of the mean scaled effect size and performed additional analyses to understand more about the responsiveness of these biomarkers.
Step 1: data curation. This initial step involved gathering a diverse array of whole blood DNAm datasets from multiple longevity-focused clinical trials and studies from various public (GEO and EMBL) and private (TruDiagnostic) repositories and standardizing their study-level metadata. Step 2: metadata standardization. Metadata for each study were harmonized to ensure consistency across datasets. Standardized variables included age, sex, sample ID, individual ID, follow-up time from baseline draw and sample type (categorizing control versus participant). Step 3: clock calculation. Using the MethylCIPHER 2.0 package, more than 110 DNAm biomarkers were calculated for each dataset. Step 4: age residuals calculation. Age was regressed out from every biomarker in each dataset to isolate age-residualized scores. Step 5: biomarker standardization age-residualized scores. For each biomarker, in every dataset, age-residualized scores were standardized (Methods) to allow comparisons across different datasets and different biomarkers. Step 6: paired t-test analysis. Paired t-tests were performed on pre- and postintervention samples to evaluate the effects of interventions on DNAm biomarkers. This statistical analysis enabled the determination of the significance of changes in biomarkers due to interventions. Step 7: heat map and follow-up analysis to explore the responsiveness of DNAm biomarkers. The heat map shows scaled effect sizes from red (positive effect) to green (negative effect), visualizing the impact of the listed interventions on the indicate DNAm biomarkers. Figure created in BioRender; Sehgal, R. https://biorender.com/4nqnt6v (2026). HBOT, hyperbaric oxygen therapy; HBT, hyperbaric therapy; ART, antiretroviral therapy; DQ, dasatinib plus quercetin; DQF, dasatinib, quercetin and fisetin.
Assessing longevity interventions: evidence for decreased epigenetic age across translAGE
To verify the compilation of a longevity interventions database—TranslAGE18—we analyzed whether 51 interventions consistently decreased epigenetic age. Our focus was on 16 prominent epigenetic clocks that were designed to be broadly applicable and not specific to any single aspect of aging. These clocks included generation 1 biomarkers (Horvath1, Horvath2 and Hannum), generation 2+ biomarkers (PhenoAge, GrimAgeV1 and DunedinPoAm38), updated clocks from earlier generations (PCHorvath1, PCHorvath2, PCHannum, PCPhenoAge, PCGrimAge, GrimAgeV2 and DunedinPACE) and generation X (Explainable) clocks developed during initial testing phases (SystemsAge, OMICmAge and DNAmEMRAge). To ensure the observed decreases in epigenetic age were genuine and not coincidental, we analyzed four control datasets and nine pro-aging event datasets (Fig. 1a and Supplementary Table 1). Two statistical analyses were performed: a one-sample t-test to test for deviations from zero effect size and paired t-tests for direct group comparisons. The results revealed that only aging events and interventions showed significant effects, with effect sizes of 0.14 (increase) and −0.06 (decrease) in the one-sample t-test (P < 0.0001). In addition, all paired t-test comparisons yielded significant results (P < 0.001).
Certain longevity intervention categories elicit stronger responses from DNAm biomarkers
We categorized the interventions in our database into four groups to assess which types had greater effects on DNAm biomarkers: lifestyle (including diet and exercise), pharmacological (for example, metformin, rapamycin, semaglutide, ketamine and anti-TNF therapy), supplements (for example, omega-3 fatty acids and folate) and medical procedures (for example, hyperbaric oxygen therapy, organ transplants and gene therapy). For the first analysis, we counted the number of significant biomarkers (out of the 16 previously identified) for each intervention and plotted the results by category on a scatter plot (Extended Data Fig. 1a and Supplementary Table 2). A paired t-test revealed that the pharmacological category had significantly more biomarkers showing effects compared with any other category, suggesting that pharmacological interventions may elicit stronger responses from DNAm biomarkers. We further conducted a one-sample t-test on the effect sizes of all clocks within each intervention category (Fig. 3b and Supplementary Table 3). This analysis showed that only Pharmacological (mean effect size of −0.09307 and t-value of 4.948) and lifestyle interventions (mean effect size of −0.0393 and t-value of 7.579) significantly reduced epigenetic age, with pharmacological interventions demonstrating a larger effect. When comparing categories, only pharmacological interventions had significantly larger effect sizes compared with others (lifestyle P = 0.0063; supplements P = 0.0009; medical procedures P = 0.0517). Together, these findings suggest that pharmacological interventions may drive a more substantial effect on DNAm biomarkers than other types of interventions.
a, A dot plot comparing the effect sizes for aging events, controls and longevity interventions across 16 biomarkers. (*P < 0.05, **P < 0.01, ***P < 0.001). Dotted line represents zero effect size line. Only significant P values are shown between different categories. b, A dot plot comparing the effect sizes of different intervention types across all 16 biomarkers. (*P < 0.05, **P < 0.01, ***P < 0.001). Dotted line represents zero effect size line. c, A heat map of biomarker responses to different interventions. The x axis lists the 16 biomarkers, and the y axis lists the interventions. Red indicates an increase in epigenetic age, green indicates a decrease in epigenetic age and gray indicates no significant effect. d, A plot comparing effect sizes across different interventions. The y axis represents the effect size, whereas the x axis represents different interventions (*P < 0.05, **P < 0.01, ***P < 0.001, # significant after multiple-testing corrections). The dotted line represents the zero effect size line. Only significant P values are shown between different categories of interventions. e, A heat map of interventions for which the TranslAGE database contains multiple studies of the same intervention, allowing consistency between studies to be examined. Red indicates an increase in epigenetic age, green indicates a decrease in epigenetic age and gray indicates no significant effect. HBOT, hyperbaric oxygen therapy; HBT, hyperbaric therapy; ART, antiretroviral therapy; DQ, dasatinib plus quercetin; DQF, dasatinib, quercetin and fisetin.
Longevity interventions have varying effects on DNAm biomarkers
Finally, we plotted the effect size of all 51 interventions for the 16 DNAm biomarkers (Fig. 3c and Supplementary Table 4) with the nonsignificant effect sizes grayed out (P value < 0.05). Several interventions seemed to affect more DNAm biomarkers (positively or negatively) than others. To quantify this without making any initial assumptions about which biomarkers are more important, we plotted all effect sizes mapping to the 16 different biomarkers for an intervention and then performed a one-way t-test on effect sizes of all interventions checking for deviations from zero effect size (Fig. 3d and Supplementary Table 5). In total, 19 interventions significantly decreased epigenetic age across the 16 biomarkers (13 after multiple-testing correction), 5 significantly increased epigenetic age across the 16 biomarkers (3 after multiple-testing correction) and the remaining 26 had no significant effects.
Consistency between biomarkers and studies supports bona fide effects of longevity interventions
Our compiled database includes a number of replicate studies of the same intervention, as well as multiple clocks that measure related constructs (for example, multiple mortality clocks). We reasoned this would be useful for determining if an intervention is indeed modifying DNAm biomarkers in a consistent manner which would support it as an effective longevity intervention. Although false positives for a given clock/intervention combination are likely to occur, it is unlikely that multiple clocks and multiple studies would repeatedly yield the same false positives. Thus we propose that interventions with bona fide effects of aging biomarkers should fulfill two rules: first, the intervention should modify DNAm biomarkers of a given generation to the same magnitude and direction in a particular study (rule 1), and a second study of the same intervention should modify the same biomarkers (rule 2). In our database, we can identify interventions that abide by both rules, one of the rules or neither of the rules (Fig. 3e). Anti-TNF therapies in patients from multiple studies (arthritis and IBD) modify nearly all generation 2+ biomarkers (both reliable and otherwise) to very similar magnitude, fulfilling both rules and supporting anti-TNF therapies as a potential method to prevent pathological aging in individuals with autoimmune disorders. Similarly, two different types of Mediterranean diets, performed in healthy cohorts, decrease similar subsets of generation 2+ DNAm biomarkers, satisfying both rule 1 and rule 2. By contrast, we see an example of two SRW supplement studies which change the same biomarkers (satisfying rule 2), but there is no consistent change in direction across those biomarkers (not satisfying rule 1—PCHorvath1 and PhenoAge decrease, whereas PCHorvath2, PCHannum and DNAmEMRAge increase). This inconsistency between biomarkers generates uncertainty about the effects of such an intervention on aging. Similarly, the five different senolytics studies have multiple biomarkers changing in different direction within a study (GrimAgeV1 and V2 versus others, not satisfying rule 1) and sometimes the same biomarker changing in opposite direction across studies (PCHorvath2, not satisfying rule 2). This example suggests that the interventional effects of senolytics, at least from an epigenetic aging perspective, may be inconsistent.
Reliable generation 2+ biomarkers demonstrate the highest responsiveness
In addition to comparing interventions, we aimed to compare the responsiveness of the 16 DNAm biomarkers discussed previously. As an initial survey, we counted the number of interventions for which each DNAm biomarker decreased or increased significantly and compared these on a scatter plot (Fig. 4a and Supplementary Table 6). Most generation 2+ reliable biomarkers such as SystemsAge, GrimAgeV2, PCGrimAge, PCPhenoAge and DunedinPACE, decreased significantly in multiple interventions and increased in none or few interventions. Notably, the only generation 2+ biomarkers that increased in more than one study were GrimAgeV1 and DunedinPoAm38, which supports utilizing their more reliable counterparts instead. Meanwhile, most generation 1 biomarkers, such as Horvath1, Horvath2 and Hannum, sporadically increased or decreased in a handful of different interventions, lacking any clear trend. DNAmEMRAge and OMICmAge changed in many interventions, though interestingly, they were a mix of increases and decreases, which may reflect the complexity of the multi-omic and phenotypic data on which they were trained. Among all the biomarkers, DunedinPACE was the most responsive to a deflection in biological age, significantly decreasing in 16 interventions and increasing in only 1. To further quantify responsiveness, we performed a second analysis where we plotted all the effect sizes for a single clock from all interventions and performed a one-way t-test checking for deviations from zero for each clock. We reasoned that because all interventions in our dataset are targeted at ameliorating pathological aging potentially by reducing biological age, then if the DNAm biomarkers are indeed responsive they should show a significant decrease when all effect sizes were pooled together (Fig. 4b and Supplementary Table 7). This analysis revealed that all generation 2+ reliable biomarkers showed significant decreases. These include GrimAgeV2 (mean of −0.0814, P = 0.0075), PCPhenoAge (mean = -0.08876, P = 0.003), PCGrimAge (mean of −0.07843, P = 0.0003), SystemsAge (mean of −0.05859, P = 0.0043) and DunedinPACE which had the largest decrease (mean of −0.0891, P = 0.0012) (significant after multiple-testing correction). By contrast, only a single generation 1 (reliable) clock significantly decreased across all interventions (PCHorvath1 mean of −0.06674, P = 0.029), and a single generation 2 nonreliable DNAm biomarker decreased (PhenoAge mean of −0.06735, P= 0.0408) (not significant after multiple-testing correction). We also regressed out DNAm imputed cell fractions and observed that nearly all biomarkers that were responsive before cell fraction regression remained responsive even after accounting for cell fractions (Extended Data Fig. 1b). Overall, both analyses converged on the same insight: reliable second-generation biomarkers are responsive to interventions, with DunedinPACE showing the largest effect size and PCGrimAge exhibiting the strongest statistical significance. This suggests potentially two criteria for a biomarker to be responsive: (1) it needs to be reliable, and (2) it needs to be a mortality or rate of aging predictor.
a, A scatter plot showing the number of interventions for which each DNAm biomarker decreases (x axis) or increases (y axis). b, A scatter plot of effect sizes of interventions on DNAm biomarkers in which each point represents one intervention study: Significant effect sizes are marked with red asterisks and blue hashtags (*P < 0.05, **P < 0.01, ***P < 0.001, # significant after multiple-testing corrections), whereas solid lines indicate the mean effect size. c, A bar plot showing agreement and disagreement among DNAm biomarkers. For each clock, the number of interventions leading to significant changes were counted (denoted by the numbers at the top). Significant changes for one clock were then categorized as either being in agreement with >50% of other clocks (blue), being in disagreement with >50% (red) or being the only clock to show a significant change (green) d, A scatter plot of the number of lifestyle interventions for which each DNAm biomarker decreases (x axis) or increases (y axis). e, A categorical scatter plot of effect sizes of lifestyle interventions on DNAm biomarkers where each point represents one intervention study. Significant effect sizes are marked with red asterisks (*P < 0.05, **P < 0.01, ***P < 0.001, # significant after multiple-testing corrections); solid lines indicate the mean effect size. f, A scatter plot of the number of pharmacological interventions for which each DNAm biomarker decreases (x axis) or increases (y axis). g, A categorical scatter plot of effect sizes of pharmacological interventions on DNAm biomarkers where each point represents one intervention study. Significant effect sizes are marked with red asterisks (*P < 0.05, **P < 0.01, ***P < 0.001, # significant after multiple-testing corrections); solid lines indicate the mean effect size.
Reliable generation 2+ biomarkers align most with other biomarkers
When selecting epigenetic clocks to evaluate intervention responses, one may want to utilize a clock that is likely to agree with other clocks on the intervention effect. If the clock were to disagree with others, then the result would be confusing to interpret, or there would be a concern that the effect was a sporadic significant result (false positive). Thus we quantified the concordance of the 16 DNAm biomarkers—the likelihood that if a DNAm biomarker detected a significant effect then others would agree on the effect—analyzed across all interventions. Thus, for each intervention where at least one biomarker responded significantly, we compared each significant biomarker with the other significant biomarkers in that study. If the biomarker agreed with greater than 50% of the remaining biomarkers, we categorized this as an agreement, if less than 50%, then a disagreement, and if it was the only significant biomarker, then as a third category of being ‘only significant’. Finally, for each of the 16 biomarkers, we divided the number in each category by the total number of significant results for that clock. We observed two trends (Fig. 4c and Supplementary Table 8). First, generation 1 biomarkers had less agreement with other biomarkers (Horvath1 of 0.66, Horvath2 of 0.5 and Hannum of 0.55) as compared with their generation 2+ counterparts (DNAmEMRAge of 0.75, OMICmAge of 0.75, GrimAgeV1 of 0.63, PhenoAge of 0.83 and DunedinPoAm38 of 0.77). Second, reliable versions showed greater agreement compared with their original versions (Horvath2 0.5 versus PCHorvath2 0.77; Hannum 0.55 versus PCHannum 0.9; GrimAge 0.6 versus PCGrimAge 0.9; PhenoAge 0.83 versus PCPhenoAge 1; and DunedinPoAm38 0.76 versus DunedinPACE 0.82) with one exception (Horvath1 0.66 versus PCHorvath1 0.44). Of particular interest were cases where a given clock was the only one that showed any significant change. Most biomarkers that showed at least one result where they were the ‘only significant’ clock also had many other results where they had large disagreements with other clocks. We interpret these clocks as having a high likelihood of false positives. However, there is one notable exception: DunedinPACE, which was unique in that it never disagreed with other clocks, yet it was the only significant clock in 17.6% of the interventions where it was significant. Although it is still possible these ‘only significant’ clock results for DunedinPACE are false positives, it is also possible that DunedinPACE is simply the most sensitive clock to intervention effects or can detect unique intervention effects.
Different DNAm biomarkers are sensitive to different intervention categories
We further wanted to test whether different DNAm biomarkers are responsive to different categories of interventions. To test this, we replicated our analysis from Fig. 4a,b, but this time limited to lifestyle and pharmacological interventions analyzed separately. In lifestyle interventions (Fig. 4d and Supplementary Table 9), DunedinPACE reported significant decreases in 8 out of the 15 interventions followed by SystemsAge (6), DunedinPoAm38 (5), DNAmEMRAge (4) and PCGrimAge (3). One-way t-tests revealed (Fig. 4e and Supplementary Table 10) that DunedinPoAm38 reported the largest decrease in effect size (mean of −0.1048, P = 0.0001), followed by DunedinPACE (−0.08887, 0.0037) and GrimAgeV2 (−0.07904, 0.0004) (all significant after multiple-testing corrections).
In comparison, after pharmacological interventions (Fig. 4f and Supplementary Table 11), GrimAgeV2 reported the highest number of significant decreases (8 out of 14) followed by PCGrimAge and OMICmAge (6). The top 4 biomarkers with largest decrease in effect size in the one-way t-test analysis (Fig. 4g and Supplementary Table 12) were DunedinPoAm38 (mean of −0.2394, P = 0.0442), GrimAgeV1 (−0.2332, 0.0151), GrimAgeV2 (−0.2092, 0.0058) and PCGrimAge (−0.1575, 0.0095) (only GrimAgeV2 and PCGrimAge significant after multiple-testing correction). Although several biomarkers did overlap between the top categories, many biomarkers were sensitive to the intervention type and their magnitude of effect size differed between the two interventional categories. Most notably, DunedinPACE was altered by fewer pharmacological interventions than other second-generation reliable biomarkers, despite its high responsiveness in lifestyle (Fig. 4d) and in uncategorized interventions (Fig. 4a). Meanwhile, GrimAgeV2 was the most sensitive to pharmacological interventions, but only changed in a single lifestyle intervention. Thus, different biomarkers were most sensitive to different intervention categories.
Study population health status is critical to DNAm biomarker response
Given the heterogeneity across interventional studies, we examined whether clinical covariates influenced DNAm biomarker responsiveness. Using linear regression, we assessed associations between DNAm effect sizes and eight study-level variables: population health status (healthy versus disease), sample size, intervention duration, mean age, age s.d., minimum age, age range and percent female (Fig. 5a and Supplementary Table 13). Population health status showed the strongest association with highly responsive biomarkers (PCPhenoAge Z of 4.83; PCGrimAge of 4.07; SystemsAge of 4.65).
a, A heat map showing standardized association strengths (Z-scores) from a linear model relating study characteristics (across all interventions) to DNAm biomarker effect sizes. Red indicates that a study characteristic is associated with higher effect sizes, green indicates lower effect sizes and gray indicates no significant association. ‘Healthy’ refers to studies conducted in participants without a diagnosed disease condition and serves as the reference group. ‘Disease’ refers to studies involving participants with a specified clinical diagnosis or pathological condition. In this coding framework, the healthy group is the baseline; therefore, positive (red) values indicate that studies conducted in disease populations are associated with larger DNAm effect sizes compared with healthy populations. min. age, minimum participant age reported in the study; age s.d., standard deviation of participant age within the study. b,c, Categorical scatter plots illustrating DNAm biomarker effect sizes in studies conducted in healthy individuals (b) and individuals with disease (c). Significant effects are marked with red asterisks (*P < 0.05, **P < 0.01, ***P < 0.001, # indicates significance after multiple-testing correction). d–i, Categorical scatter plots comparing effect sizes between healthy and disease populations for specific DNAm biomarkers: PCPhenoAge (d), SystemsAge (e), PCGrimAge (f), GrimAgeV2 (g), PCHannum (h) and DunedinPACE (i). NS, not significant.
We then tested whether DNAm biomarkers responded differently in healthy versus disease populations. One-sample t-tests revealed that (Fig. 5b,c and Supplementary Tables 14 and 15) several biomarkers were significantly responsive in disease groups (for example, PCPhenoAge mean of −0.32, P = 0.0057; PCGrimAge of −0.22, P = 0.0022; SystemsAge of −0.213, P = 0.0105; all significant after multiple-testing correction), whereas only a few showed responsiveness in healthy groups (for example, DunedinPACE of −0.074, P = 0.0143; PCGrimAge of −0.043, P = 0.0191; none were significant after multiple-testing correction).
To determine whether this pattern reflected a few disease studies (n = 10) versus the larger number of healthy studies (n = 41), we compared biomarker effect sizes between groups using unpaired t-tests (Fig. 5d–i and Supplementary Table 16). Several biomarkers, including PCPhenoAge (mean difference of 0.28, P < 0.0001), SystemsAge (mean difference of 0.21, P < 0.0001) and GrimAgeV2 (mean difference of 0.14, P = 0.02), showed significantly larger decreases in disease populations. Even PCGrimAge, which was responsive in both groups, showed a greater response in disease studies (mean difference of 0.18, P = 0.002). Only DunedinPACE responded similarly in both populations (mean difference of 0.07, P = 0.25), suggesting it may be the most robust across varying health contexts.
GenX DNAm biomarkers offer mechanistic insights into responsiveness
Clocks such as SystemsAge, GrimAge and OMICmAge were built from component DNAm biomarkers to better explain the specific aging changes that they are capturing. Thus, we also tested responsiveness for 78 generation explainable (GenX) DNAm biomarkers (Supplementary Table 17). We observed 39 biomarkers with significant changes across all interventions (Fig. 6a). Examples of biomarkers with particularly high responsiveness included PCGrimAge components PCPACKYRS (effect size of −0.05824, t = 4.636), PCCystatinC (−0.1404, 3.630) and PCTIMP1 (−0.08526, 3.666); SystemsAge components kidney (−0.1021, 3.753), lung (−0.08812, 4.615) and musculoskeletal (−0.07132, 3.729); and finally, OMICmAge components EBP-N-acetyl-isoputreanine (−0.09316, 3.506), EBP-ureidopropionic acid (−0.05658, 3.733), EBP-Mimecan (−0.1238, 4.047), EBP-Cystine (−0.1162, 3.897), EBP-CBPB2 (0.06373, 3.497) and EBP-BMP1 (0.09398, 4.183).
a, A scatter plot showing intervention effect sizes for GenX DNAm biomarkers. The x axis lists individual GenX biomarkers, and the y axis shows the corresponding standardized effect sizes. Each point represents the estimated effect size for a biomarker across interventions. Horizontal dashed lines indicate significance thresholds. Red points and red asterisks denote significant effects (*P < 0.05, **P < 0.01, ***P < 0.001), whereas gray points indicate nonsignificant effects. P values test whether the intervention-associated effect size differs significantly from zero. b, A heat map of whole-body aging DNAm biomarkers (y axis) across interventions (x axis). Red indicates an increase in epigenetic age, green indicates a decrease in epigenetic age and gray indicates no significant effect. c, A heat map of system-specific DNAm biomarkers (y axis) across interventions (x axis). The color scale is the same as in b. d, A heat map of scaled effect sizes for organ-system-specific DNAm biomarkers across dietary interventions. The color scale is consistent with b and c. e, A heat map of scaled effect sizes for DNAm-based metabolite and protein proxies across dietary interventions. The color scale is consistent with b–d. f, A scatter plot comparing intervention effect sizes estimated by StocP (x axis) and PhenoAge (y axis). Each point represents a single intervention. Red points indicate interventions associated with significant increases in epigenetic age across the 16 aging biomarkers, green points indicate interventions associated with significant decreases in epigenetic age and gray points indicate interventions with no significant overall effect. g, A scatter plot comparing intervention effect sizes estimated by StocH (x axis) and Horvath1 (y axis). Each point represents a single intervention and is colored according to its overall effect on epigenetic age as described for f. h, A scatter plot comparing intervention effect sizes estimated by AdaptAge (x axis) and DamAge (y axis). Each point represents a single intervention and is colored according to its overall effect on epigenetic age as described for f.
To understand how GenX clocks can provide more insight into responsiveness, we examined specific intervention, selected on the basis of whether they fell into two categories: (1) significantly decreasing epigenetic age in only a handful of generation 2 reliable clocks (smoking cessation—SystemsAge; umbilical cord blood transfusion—DunedinPACE and GrimAgeV2; metformin—SystemsAge, PCGrimAge and PCPhenoAge; hyperbaric oxygen therapy—DunedinPACE; Vegan Diet—DunedinPACE and SystemsAge; TruLacta supplement—DunedinPACE) or (2) not significantly decreasing epigenetic age in any generation 2 reliable clocks (gastric bypass) (Fig. 6b and Supplementary Table 18). The sporadic responsiveness of clocks to the first category of interventions may suggest they are false positives, whereas the lack of responsiveness in the second category may suggest they do not modify epigenetic age. However, we considered the possibility that these interventions simply had very specific effects that may not be detectable using a general aging clock. Accordingly, the SystemsAge components suggested these interventions may be impacting different systems (Fig. 6c, Extended Data Fig. 2 and Supplementary Table 18). For example, the two interventions that did not modify any general Gen 2 reliable clock do show significant decreases in epigenetic age of specific systems (gastric bypass decreases metabolic score with effect size 0.43). Other interventions include smoking cessation seeing maximal decrease in lung (0.20), umbilical cord transfusion in musculoskeletal (0.13) and hyperbaric oxygen therapy in lung (0.27). Although other systems saw multi-system decreases such as metformin with inflammation (0.80), brain (0.70) and metabolic (0.70) decreasing the most. In the vegan diet, the largest decrease was seen in inflammation (0.23) and musculoskeletal (0.17). However, in TruLacta supplement, kidney (0.29), immune (0.28) and inflammation (0.24) reported the greatest impact after the intervention.
Finally, given we had a multitude of diets (seven) in our study (low-carb diet, low-fat diet, green mediterranean diet, healthy guidelines diet, red meat mediterranean diet, CR/IMF (calorie restriction/intermittent fasting) and vegan diet) we wished to understand if GenX DNAm biomarkers could provide deeper mechanistic understanding specifically into dietary interventions. We looked at the 11 system scores and specific metabolic related protein and metabolite epigenetic proxy biomarkers (Fig. 6d,e and Supplementary Table 18). Among the 11 system-specific biomarkers, musculoskeletal was the only score that significantly decreased across all seven diets. Meanwhile, two system scores (metabolic and inflammation) were responsive in six of the seven diets. When looking at the protein and metabolite proxies, we observed the epigenetic proxy of triglycerides to be significant across six of the seven diets, whereas the glucose proxy was significant in four of seven diets, suggesting great potential for these epigenetic proxies in measuring diet-based interventions.
Responsiveness of stochastic and causal components
Recent advances have led to a new generation of DNAm biomarkers that aim to separate meaningful biological signals from random noise19,20. These biomarkers appear to contain both stochastic (random) and nonstochastic (predictable, biologically driven) components. These studies found that first-generation clocks such as Horvath1, which estimate chronological age, mostly reflect stochastic epigenetic noise. By contrast, second-generation clocks (generation 2+) such as PhenoAge and GrimAge are believed to capture more biologically meaningful, nonrandom aging signals. However, those findings were based on cross-sectional data. To test this more rigorously, we used the longitudinal datasets in TranslAGE18.
We examined how much of the change in each clock could be attributed to stochastic versus nonstochastic components. For example, we found that Horvath1 was strongly correlated with its stochastic component (R = 0.50), whereas PhenoAge showed almost no correlation with its own stochastic component (R = −0.06), suggesting that PhenoAge reflects more structured biological aging (Fig. 6f,g).
Clocks have also been developed to isolate causal DNAm changes that may reflect either harmful or adaptive aging processes16. We revisited the DamAge–AdaptAge framework and found a strong inverse relationship between their changes (R = −0.89). Yet, most interventions reduced both components, indicating an overall slowdown of biological change rather than a shift from harmful to adaptive processes.
