LinAge2: Providing actionable insights and benchmarking with epigenetic clocks

24 Biological aging is marked by a decline in resilience at the cellular and systemic levels, driving an 25 exponential increase in mortality risk. Here, we evaluate several clinical and epigenetic clocks for their 26 ability to predict mortality, demonstrating that clocks trained on survival and functional aging 27 outperform those trained on chronological age. We present an enhanced clinical clock that predicts 28 mortality more accurately and provides actionable insights for guiding personalized interventions. 29 These findings highlight the potential of mortality-predicting clocks to inform clinical decision-making 30 and promote strategies for healthy longevity. 31 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 3 Main 32 Biological aging is characterized by the progressive decline in intrinsic biological resilience that is 33 associated with an exponential increase in mortality, expressed in the demographic “Gompertz 34 mortality law”1. Not all humans age at the same rate since genetics, lifestyle and stochastic factors 35 can affect future mortality and morbidity trajectories. Consequently, individual true biological age (BA) 36 is not identical to calendar or chronological age (CA). The true BA of an individual can be uniquely 37 defined as the age at which subjects of a reference cohort have the same risk of age-dependent 38 disease and all-cause mortality as the subject in question. Tools to accurately track changes in true 39 BA are essential for the development and validation of novel life- and healthspan-optimizing diet, 40 lifestyle, supplement and drug interventions. 41 42 Biological aging “clocks” are computational tools that estimate individual true BA based on 43 demographic, clinical, and/or molecular data. CA itself is widely used for both clinical prognostication 44 and decision-making, and can be viewed as a first order approximation of true BA. The ideal BA clock 45 should predict individual Gompertz mortality risk with higher accuracy than CA. Some aging clocks, 46 including most clinical clocks, explicitly include CA as a covariate, using biological features to estimate 47 a correction factor aimed at providing a better estimate of true BA. CA in this case is used as a proxy 48 for effects and mechanisms, such as entropic damage, not captured by the clock itself. Of course, the 49 ideal clock would include all relevant processes, wherein the model would assign zero or negligible 50 weight to CA. 51 52 Aging clocks are generalizations of current clinical risk markers that predict disease-specific morbidity 53 and, in some cases, mortality. Aging clocks should similarly enable early detection of hidden or 54 subclinical diseases, surpassing the capabilities of diagnostics by identifying disease processes years 55 or decades before overt disease is present. Secondly, to inform risk-to-benefit estimates (clinical 56 equipoise), aging clocks should capture all-cause mortality holistically, providing value beyond organ 57 or disease-specific risks. Thirdly, aging clocks must be sensitive to individual variations in biological 58 resilience. Finally, aging clocks should provide tools for mechanistic interpretation and provide 59 actionable insights, facilitating targeted interventions. To date, none of the existing clocks meet all 60 these criteria. 61 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 4 62 To evaluate the performance of aging clocks, we can compare them to a hypothetical “ideal” clock, 63 which we term “CrystalAge”. This optimal clock would predict disease-specific and all-cause mortality 64 at the individual level with near-perfect accuracy, essentially forecasting an individual’s date of death 65 (Fig. 1a and d). While practically impossible, in retrospective studies, we can determine the 66 theoretically optimal performance of CrystalAge and use it as a benchmark to evaluate the 67 performance of existing aging clocks. 68 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 5 69 Fig. 1: LinAge2 predicts 20-year all-cause mortality and tracks with healthspan markers. a-c, 70 Kaplan-Meier survival curves showing 20-year survival in the 65-74 CA bin ( n=631). For each clock, 71 subjects were stratified by selecting the lowest (best, solid line) and highest (worst, dotted line) 25% 72 quartiles for BA. Clocks within the same quartile were compared using log-rank tests with Benjamini-73 Hochberg correction. Areas shaded indicate 95% error bands for lines of the same color. b, 74 Compared to ChronAge, use of LinAge2 BA results in a significant survival difference for the lowest 75 25% BA quartile ( P=6.16E-04), but not for the highest 25% quartile ( P=0.07). PhenoAge Clinical did 76 not significantly outperform ChronAge in predicting survival in this age bin. c, LinAge2 significantly 77 outperformed DunedinPoAm (P=1.09E-02) and PhenoAge DNAm (P=1.37E-03) in the lowest 25% BA 78 quartile, but not GrimAge2 ( P=0.22). In the highest 25% quartile, while LinAge2 significantly 79 outperformed PhenoAge DNAm ( P=0.03), the differences between LinAge2 and DunedinPoAm 80 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 6 (P=0.11) and GrimAge2 (P=0.58) did not reach statistical significance. e, ROC analysis revealed that 81 LinAge2 (area under the curve (AUC)=0.8684) was significantly more informative than PhenoAge 82 Clinical (AUC=0.8479, P=6.35E-05) and ChronAge (AUC=0.8288, P=3.16E-10) in predicting future 83 mortality ( n=2,036). LinAge2 performed similarly to LinAge (AUC=0.8647). f, LinAge2 also 84 outperformed PhenoAge DNAm (AUC=0.7859, P=4.44E-07) and GrimAge2 (AUC=0.8233, P=0.02) in 85 predicting 20-year mortality ( n=1,065). Although GrimAge2 outperformed ChronAge (AUC=0.7933, 86 P=2.74E-03) in predicting 20-year mortality, PhenoAge DNAm did not ( P=0.47). a,d, HorvathAge, 87 HannumAge, and ChronAge did not significantly differ in predicting mortality risk (AUCs=0.7776, 88 0.7978 and 0.7933, respectively, n=1,065). ROC curves were compared using DeLong’s test. 89 a,b,d,e,f, CrystalAge, a theoretical perfect clock shown for reference, accurately identifies individuals 90 at risk of dying (AUC=1), whereas RandomAge adds random gaussian noise of +10 years to CA. g-k, 91 Violin plots for each clock categorized into low (biologically younger/best 25% quartile) and high 92 (biologically older/worst 25% quartile) groups plotted against healthspan markers: cognitive scores 93 (digit symbol substitution test), gait speed, ability to work, and ability to perform all instrumental and 94 basic activities of daily living (iADLs and bADLs). Groups (BA high versus low) were compared using 95 two-sided t-tests. Median value, lower (25 th) and upper (75 th) percentiles are indicated. Lines extend 96 to +1.5 times interquartile range, with points outside this range drawn individually. The violin shape 97 indicates the probability density function. yo, years old. HA, HorvathAge. LA2, LinAge2. GA2, 98 GrimAge2. DPA, DunedinPoAm. 99 100 Taking inspiration from Levine’s PhenoAge clinical clock2, we recently developed and validated clinical 101 aging clocks (PCAge, LinAge) based on linear dimensionality reduction by matrix factorization 102 (singular value decomposition) and demonstrated them to be highly predictive in terms of future 103 disease-specific and all-cause mortality3. These clocks have since been applied in a range of clinical 104 settings and, taking advantage of user feedback, we have implemented several improvements, 105 creating an updated version of these clocks (LinAge2). Like LinAge, we trained LinAge2 in the 106 National Health and Nutrition Examination Survey (NHANES) IV 1999-2000 wave before testing it in 107 the 2001-2002 wave. LinAge2 further reduces the number of rarely measured parameters and 108 emphasizes interpretability. For a detailed description of LinAge2's features and construction, refer to 109 Methods. 110 111 Many aging clocks have been developed, with epigenetic or DNA methylation (DNAm) clocks most 112 widely recognized and well-established. Several epigenetic clocks have been commercially licensed 113 for applications, including estimating CA (HorvathAge 4, HannumAge 5), optimizing life insurance 114 policies (PhenoAge DNAm6, GrimAge7), and monitoring the rate of aging (DunedinPoAm8)9. Recently, 115 a dataset of pre-calculated epigenetic clock ages has been published for the NHANES 1999-2002 116 waves, permitting direct comparison of the predictive power of CA, the original LinAge, LinAge2 and 117 PhenoAge clinical clocks, and the HorvathAge, HannumAge, PhenoAge DNAm, GrimAge2 10 and 118 DunedinPoAm epigenetic clocks. 119 120 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 7 To compare efficacy in predicting mortality, we performed survival and receiver operating 121 characteristic (ROC) analyses on 20- and 10-year mortality in the NHANES 2001-2002 test cohort. 122 Compared to CA, LinAge2 demonstrated significant survival differences across all age bins, whereas 123 PhenoAge Clinical did not (Fig. 1b, Extended Data Fig. 1b and e). LinAge2 performed similarly to 124 LinAge and demonstrated superior predictive power for future mortality compared to PhenoAge 125 Clinical and CA (Fig. 1e and Extended Data Fig. 2b). Surprisingly, LinAge2 also outperformed 126 PhenoAge DNAm and DunedinPoAm in predicting age-specific survival differences (Fig. 1c, Extended 127 Data Fig. 1c and f) and future mortality (Fig. 1f and Extended Data Fig. 2c). In contrast, PhenoAge 128 DNAm, HorvathAge, and HannumAge did not significantly differ from CA in predicting future mortality 129 (Fig. 1a, d and f, Extended Data Fig. 1a and d, and Extended Data Fig. 2a and c). LinAge2 and 130 GrimAge2 performed similarly in predicting future mortality (Fig. 1f and Extended Data Fig. 2c) and 131 survival across all age bins (Fig. 1c, Extended Data Fig. 1c and f). 132 133 While mortality prediction is an important function of aging clocks, it is important to evaluate if clock 134 ages are similarly predictive of functional status and healthspan. We tested this for the same clinical 135 and epigenetic clocks by comparing markers of functional and health status in individuals selected by 136 each clock to be in the lowest 25% BA quartile (biologically younger, low) with those in the highest 137 25% quartile (biologically older, high). Our analysis revealed that LinAge2 low was associated with 138 superior healthspan markers, including higher cognitive scores, faster gait speed, ability to work, and 139 performance of all instrumental and basic activities of daily living (iADLs and bADLs) (Fig. 1g-k). 140 Conversely, individuals in the LinAge2 high group had poorer healthspan, with statistically significant 141 differences between the two groups across all markers (Fig. 1g-k). Similar trends were observed for 142 GrimAge2 and DunedinPoAm, with statistically significant differences between the low and high 143 groups across most healthspan markers, except for the ability to perform all bADLs (Fig. 1g-k and 144 Extended Data Fig. 3). In contrast, no statistically significant differences were found between 145 HorvathAge low and HorvathAge high across healthspan markers (Fig. 1g-k). Our findings on 146 healthspan markers and mortality, for HorvathAge, HannumAge, PhenoAge DNAm and GrimAge2, 147 corroborate similar findings for 10-year survival in 490 subjects of the Irish Longitudinal Study on 148 Aging11. 149 150 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 8 A significant drawback of many existing aging clocks is that they lack interpretability and actionable 151 insights, making it challenging to develop targeted interventions. However, one key benefit of clinical 152 clocks is that they are built from parameters directly related to the underlying disease mechanisms, 153 enabling easier interpretation of clock residuals and providing actionable insights into disease 154 pathophysiology. Individual age-associated principal components (PCs) identify clusters of features 155 that change in a coordinated manner during aging. Analyzing individual PCs can provide valuable 156 insights into the underlying patterns and trajectories of aging-related changes. Using heatmaps, we 157 visualized the predictive power of individual PCs relative to clinical outcomes, including sex-specific 158 causes of death and chronic diseases (Fig. 2). Supplementary Table 5 provides detailed insights into 159 the interpretation of each PC, including associations with causes of death, chronic diseases, lifestyle 160 factors, and potential aging mechanisms, as well as suggested interventions to optimize each PC and 161 lower BA. 162 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 9 163 Fig. 2: Heatmaps illustrating the associations between clinical outcomes and PCs analyzed 164 using multivariate logistic regression. Associations of PCs with specific causes of death at a-b, 10-165 20 year and c-d, 0-5 year follow-up for male and females, respectively. Strength of association (odds 166 ratio) is represented using a red color scale. PCs that are strongly positively associated with a specific 167 cause of death are bright red. For cause of death, negative associations were truncated by setting 168 their values to zero (i.e. no harm). e-f, Association of PCs with specific chronic diseases, and 169 measures of lifestyle and socioeconomic status. Strength of association is represented using a blue-170 red scale ranging from -1 (blue, negative association) to 2 (red, positive association). PCs that are 171 strongly positively associated (positive risk ratios) with a specific disease are bright red, whereas PCs 172 that are strongly negatively associated (negative risk ratios) with a specific disease are bright blue. 173 174 Accurately predicting patient outcomes and allocating healthcare resources is a significant challenge 175 in clinical practice12. Currently, clinicians rely heavily on CA to make these decisions. However, here 176 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 10 we show that mortality-predicting clocks, such as LinAge2 and GrimAge2, outperform CA in predicting 177 mortality risk across timeframes, ranging from 2-20 years (Fig. 1, Extended Data Fig. 4). Moreover, 178 clinical clocks can also predict specific causes of death within a 5-year window (Fig. 2c and d). This 179 illustrates that clock-based BAs are more accurate and informative estimates of true BA than CA itself. 180 By providing a more precise metric of biological status than CA alone, BA can enable clinicians to 181 better support patients and their caregivers in navigating healthcare choices, including end-of-life 182 care. 183 184 Overall, our analysis reveals that, regardless of feature space (methylation or clinical), aging clocks 185 trained to predict mortality or functional aging outcomes provide more predictive value in terms of 186 clinical decision-making. Surprisingly, clinical aging clocks still outperform several prominent mortality-187 predicting and functional epigenetic clocks, including PhenoAge DNAm and DunedinPoAm, in 188 predicting future mortality. A key advantage of LinAge2 lies in its interpretability. Because principal 189 component analysis is a linear matrix factorization technique, the resulting model is easier to interpret 190 than nonlinear alternatives 13. Latent variables based on linear dimensionality reduction (PCs), 191 especially those based on clinical parameters, are comparatively easy to understand and interpret, 192 making them more actionable. This enables clinical aging clocks like LinAge2 to detect hidden or 193 subclinical diseases and inform primordial prevention strategies. By identifying individuals at high risk 194 of developing specific diseases, healthcare providers can implement targeted interventions early and 195 proactively. By casting specific risk in terms of BA acceleration, aging clocks can significantly increase 196 compliance and adherence with specific health recommendations14. For example, male smokers with 197 high PC5M values in LinAge2 are at increased risk of death from chronic lung disease and should be 198 screened and advised to quit smoking (Fig. 2 and Supplementary Table 5). 199 200 All current aging clocks, regardless of feature space (e.g. clinical, methylation, proteomics, etc.) and 201 target (mortality, functional outcomes, disease, or CA) share significant limitations. Most importantly, 202 many current clocks employ linear techniques (e.g. principal component analysis/singular value 203 decomposition, regression-based predictions), which limit their ability to distinguish between aging 204 signatures and those of age-dependent diseases, and to learn U-shape response patterns. Current 205 clocks therefore inherently conflate intrinsic biological aging with disease-specific signatures (hidden 206 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 11 sickness, primordial disease signatures). Nonlinear approaches including generative artificial 207 intelligence and artificial neural networks are being investigated and could offer improved models, but 208 their increased complexity pose a significant challenge for interpretation 15,16. Next generation clocks 209 will need to differentiate between disease signatures and intrinsic aging, and quantify intrinsic 210 biological resilience. Further theoretical work will be required to deconvolute these disease-centric 211 signatures from determinants of intrinsic resilience and entropic aging17-19. Advancing next generation 212 clocks is crucial to equip healthcare providers with the essential tools needed to make informed 213 decisions regarding targeted interventions that support healthy longevity in populations where 214 healthcare needs are increasingly dominated by aging. 215 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 12

216 Motivation for enhancing LinAge2 217 The original PCAge and LinAge 3 both utilized some parameters that are not routinely collected. 218 LinAge has been utilized by several clinics worldwide, and we have received informal feedback 219 regarding its use. Common suggestions for enhancing the clock include: (i) improving handling of 220 outliers and threshold effects, (ii) further refining the clinical parameters, especially removing serum 221 fibrinogen due to the need for a specialized sodium citrate tube, (iii) providing additional tools to 222 improve the interpretability of PCs, and (iv) providing specific strategies to optimize each PC to lower 223 BA. PCs can also be sensitive to outliers, thresholding and batch effects. We developed LinAge2 in 224 response to these concerns. 225 226 We followed the same workflow, as previously described 3, to construct LinAge2 but with several 227 modifications. To enhance LinAge2, we refined the clinical parameters by reducing the total number to 228 60, removing serum fibrinogen (Supplementary Table 2). We also addressed outliers and thresholding 229 by capping outliers at six standard deviations and log-transforming additional parameters 230 (Supplementary Table 2). Batch effects were mitigated through z-score normalization by median and 231 median absolute deviation to a younger, generally healthy cohort (age 40-50 years), separately for 232 males and females (Supplementary Table 2), generating sex-specific PCs. The loadings for male and 233 female PCs are provided in Supplementary Table 3, and sex-specific weights of the Cox proportional 234 hazards models are listed in Supplementary Table 4. 235 236 A parametrized version of LinAge2 is provided as previously described3 (Supplementary Table 2). The 237 baseline characteristics of the study participants are listed in Supplementary Table 1. 238 239 PhenoAge Clinical and epigenetic clocks 240 PhenoAge Clinical was implemented using the equation from the original publication. The dataset of 241 pre-calculated epigenetic clock ages published for the NHANES 1999-2002 waves were obtained 242 from https://wwwn.cdc.gov/nchs/nhanes/dnam/ and analyzed. 243 244 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 13 Construction of healthspan markers 245 The digit symbol substitution test score (NHANES variable ‘CFDRIGHT’) was used as a cognitive 246 measure. Gait speeds were obtained by taking the total distance walked (20 feet or 6.096 meters) 247 divided by the time taken (NHANES variable ‘MSXWTIME’). Differences in cognitive scores and gait 248 speeds were calculated as the percent difference between a control group (middle 50% of all 249 subjects), for younger (best 25% quartile) and older (worst 25% quartile) groups. The ability to work 250 was established using the NHANES variable ‘PFQ048’. The ability to perform all instrumental 251 activities of daily living (iADLs) was a combination of the NHANES variables ‘PFQ060A’, ‘PFQ060F’, 252 ‘PFQ060G’, ‘PFQ060Q’, PFQ060R’ and ‘PFQ060S’, while the ability to perform all basic activities of 253 daily living (bADLs) was a combination of the NHANES variables ‘PFQ060B’, ‘PFQ060C’, ‘PFQ060H’, 254 ‘PFQ060I’, ‘PFQ060J’, ‘PFQ060K’ and ‘PFQ060L’. Participants had to have either no difficulty or 255 some difficulty in all the variables to be deemed able to perform all iADLs or all bADLs. 256 257 Heatmap analysis 258 Using the ‘nnet’ 20 (version 7.3-19) R package, heatmaps were generated to evaluate the predictive 259 values for PCs included in LinAge2. For each parameter, we attempted to predict status 260 (diseased/compromised or not) using multivariate logistic regression with the clock PCs as covariates. 261 PCs that received a statistically insignificant ( P>0.05) weight in the logistic regression model were 262 assigned zero weights (white). The remaining PCs ( P<0.05) were assigned color values according to 263 their weight in the model (see Fig. 2f legend for color mapping). 264 265 Statistics and reproducibility 266 For the NHANES IV 1999-2002 waves, we excluded: participants top-coded at age 85 years, as we 267 could not ascertain the exact CAs of these adults, and participants who died from accidental deaths, 268 as these were deemed to be not age-related. 269 270 Survival analyses were performed using log-rank tests with Benjamini-Hochberg correction. ROC 271 curves were compared using DeLong’s test. For healthspan markers, two-sided t-tests were used to 272 compare between the low and high clock groups. All statistical analyses were performed using R 273 version 4.2.0 (https://www.R-project.org/). 274 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 14 275 Data availability 276 All datasets used are publicly available online at https://wwwn.cdc.gov/nchs/nhanes/Default.aspx. 277 There were no restrictions on data availability. This study was reported according to STROBE 278 guidelines for cohort studies. 279 280

281 1 Gompertz, B. On the nature of the function expressive of the law of human mortality : and on 282 a new mode of determining the value of life contingencies ; In a letter to Francis Baily / by 283 Benjamin Gompertz. (Printed by W. Nicol, 1825). 284 2 Liu, Z. et al. A new aging measure captures morbidity and mortality risk across diverse 285 subpopulations from NHANES IV: A cohort study. PLoS Med 15, e1002718 (2018). 286 https://doi.org:10.1371/journal.pmed.1002718 287 3 Fong, S. et al. Principal component-based clinical aging clocks identify signatures of healthy 288 aging and targets for clinical intervention. Nat Aging 4, 1137-1152 (2024). 289 https://doi.org:10.1038/s43587-024-00646-8 290 4 Horvath, S. DNA methylation age of human tissues and cell types. Genome Biol 14, R115 291 (2013). https://doi.org:10.1186/gb-2013-14-10-r115 292 5 Hannum, G. et al. Genome-wide methylation profiles reveal quantitative views of human 293 aging rates. Mol Cell 49, 359-367 (2013). https://doi.org:10.1016/j.molcel.2012.10.016 294 6 Levine, M. E. et al. An epigenetic biomarker of aging for lifespan and healthspan. Aging 295 (Albany NY) 10, 573-591 (2018). https://doi.org:10.18632/aging.101414 296 7 Lu, A. T. et al. DNA methylation GrimAge strongly predicts lifespan and healthspan. Aging 297 (Albany NY) 11, 303-327 (2019). https://doi.org:10.18632/aging.101684 298 8 Belsky, D. W. et al. Quantification of the pace of biological aging in humans through a blood 299 test, the DunedinPoAm DNA methylation algorithm. Elife 9 (2020). 300 https://doi.org:10.7554/eLife.54870 301 9 Moqri, M. et al. Biomarkers of aging for the identification and evaluation of longevity 302 interventions. Cell 186, 3758-3775 (2023). https://doi.org:10.1016/j.cell.2023.08.003 303 10 Lu, A. T. et al. DNA methylation GrimAge version 2. Aging (Albany NY) 14, 9484-9549 (2022). 304 https://doi.org:10.18632/aging.204434 305 11 McCrory, C. et al. GrimAge Outperforms Other Epigenetic Clocks in the Prediction of Age-306 Related Clinical Phenotypes and All-Cause Mortality. J Gerontol A Biol Sci Med Sci 76, 741-307 749 (2021). https://doi.org:10.1093/gerona/glaa286 308 12 Orlovic, M. et al. Accuracy of clinical predictions of prognosis at the end-of-life: evidence from 309 routinely collected data in urgent care records. BMC Palliat Care 22, 51 (2023). 310 https://doi.org:10.1186/s12904-023-01155-y 311 13 Sayed, N. et al. An inflammatory aging clock (iAge) based on deep learning tracks 312 multimorbidity, immunosenescence, frailty and cardiovascular aging. Nat Aging 1, 598-615 313 (2021). https://doi.org:10.1038/s43587-021-00082-y 314 14 Bonner, C. et al. Interventions Using Heart Age for Cardiovascular Disease Risk 315 Communication: Systematic Review of Psychological, Behavioral, and Clinical Effects. JMIR 316 Cardio 5, e31056 (2021). https://doi.org:10.2196/31056 317 15 Qiu, W., Chen, H., Kaeberlein, M. & Lee, S. I. ExplaiNAble BioLogical Age (ENABL Age): an 318 artificial intelligence framework for interpretable biological age. Lancet Healthy Longev 4, 319 e711-e723 (2023). https://doi.org:10.1016/S2666-7568(23)00189-7 320 16 Nusinovici, S. et al. Application of a deep-learning marker for morbidity and mortality 321 prediction derived from retinal photographs: a cohort development and validation study. 322 Lancet Healthy Longev 5, 100593 (2024). https://doi.org:10.1016/S2666-7568(24)00089-8 323 17 Denisov, K. A., Gruber, J. & Fedichev, P . O. Discovery of Thermodynamic Control Variables 324 that Independently Regulate Healthspan and Maximum Lifespan. bioRxiv, 325 2024.2012.2001.626230 (2024). https://doi.org:10.1101/2024.12.01.626230 326 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 15 18 Perevoshchikova, K. & Fedichev, P . O. Differential Responses of Dynamic and Entropic Aging 327 Factors to Longevity Interventions. bioRxiv, 2024.2002.2025.581928 (2024). 328 https://doi.org:10.1101/2024.02.25.581928 329 19 Tarkhov, A. E., Denisov, K. A. & Fedichev, P. O. Aging clocks, entropy, and the challenge of 330 age reversal. AgingBio 2, e20240031 (2024). https://doi.org:10.59368/agingbio.20240031 331 20 Venables, W. N. & Ripley, B. D. Modern Applied Statistics with S . Fourth edn, (Springer, 332 2002). 333 334

335 We thank the National Health and Nutrition Examination Survey participants and staff who made this 336 study possible. We thank C. Chen for her careful reading of this manuscript. This research was 337 funded by the Ministry of Education in Singapore, grant numbers IG21-SG007 and A-0007215-00-00, 338 to J.G. S.F. is supported by the Research Training Fellowship (MOH-001294-00) from the National 339 Medical Research Council Singapore. This work was supported by the Lien Foundation. 340 341 Author contributions 342 S.F., B.K.K. and J.G. conceived, conceptualized and designed the study. S.F., K.A.D. and J.G. 343 analyzed and interpreted the data. S.F., B.K.K. and J.G. wrote the first draft of the paper. 344 345 Competing interests 346 The authors declare no competing interests. 347 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 16 Additional information 348 349 Extended Data Fig. 1: LinAge2 predicts survival in chronologically 55-64 and 75-84 year old 350 individuals. Kaplan-Meier survival curves showing 20-year survival in the a-c, 55-64 CA bin (n=657) 351 and d-f, 75-84 CA bin ( n=348) in the test cohort. a,d, HannumAge, HorvathAge, and ChronAge 352 showed no statistically significant differences in survival. b, LinAge2 demonstrated significant survival 353 differences compared to ChronAge in both the best 25% ( P=2.01E-03) and worst 25% ( P=0.03) 354 quartiles. In contrast, PhenoAge Clinical showed a significant difference only in the best 25% quartile 355 (P=3.80E-02). c,f, In the best 25% quartile, LinAge2 outperformed DunedinPoAm ( P=6.45E-03 and 356 P=1.39E-02 in the 55-64 and 75-84 CA bins, respectively) and PhenoAge DNAm ( P=8.90E-03 and 357 P=1.73E-02 in the 55-64 and 75-84 CA bins, respectively). LinAge2 and GrimAge2 performed 358 similarly with no significant differences between them. In the worst 25% quartile, no significant 359 differences in survival were found between LinAge2, DunedinPoAm, PhenoAge DNAm, and 360 GrimAge2. e, Compared to ChronAge, both LinAge2 (P=4.56E-03) and PhenoAge Clinical (P=3.16E-361 02) showed significant survival differences in the best 25% quartile, but not in the worst 25% quartile. 362 Clocks were compared using log-rank tests with Benjamini-Hochberg correction. Areas shaded 363 indicate 95% error bands for lines of the same color. yo, years old. 364 365 366 Extended Data Fig. 2: ROC curves for 10-year all-cause mortality in the test cohort. a, There 367 were no significant differences in the AUCs between HorvathAge (AUC=0.7425), HannumAge 368 (AUC=0.7612), and ChronAge (AUC=0.7501) ( n=1,065). b, LinAge2 (AUC=0.8468) was significantly 369 more informative than PhenoAge Clinical (AUC=0.8203, P=6.91E-05) and ChronAge (AUC=0.7946, 370 P=2.65E-09) in predicting future mortality ( n=2,036). LinAge2 performed similarly to LinAge 371 (AUC=0.8383). c, Compared to LinAge2 (AUC=0.8144), PhenoAge DNAm (AUC=0.7390, P=2.84E-372 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint 17 06) and GrimAge2 (AUC=0.7801, P =1.81E-03) were significantly less predictive of 10-year follow-up 373 (n=1,065). Although GrimAge2 outperformed ChronAge (AUC=0.7501, P =0.01) in predicting 10-year 374 mortality, PhenoAge DNAm did not (P=0.39). ROC curves were compared using DeLong’s test. 375 376 377 Extended Data Fig. 3: DunedinPoAm tracks with healthspan markers. Differences between 378 DunedinPoAm low (slow aging) and DunedinPoAm high (fast aging) were significant for: a, ability to 379 work; b, ability to perform instrumental activities of daily living (iADLs); but not for c, ability to perform 380 basic activities of daily living (bADLs). Median value, lower (25 th) and upper (75 th) percentiles are 381 indicated. Lines extend to + 1.5 times interquartile range, with points outside this range drawn 382 individually. The violin shape indicates the probability density function. 383 384 385 Extended Data Fig. 4: Clinical clocks are better predictors of 2-year mortality than CA. ROC 386 curves for 2-year all-cause mortality in the test cohort. LinAge2 (AUC=0.8294, P =4.02E-06), LinAge 387 (AUC=0.8003, P=3.91E-04) and PhenoAge Clinical (AUC=0.7870, P=3.24E-03) were significantly 388 more predictive of 2-year follow-up than ChronAge (AUC=0.7120) ( n=2,036). ROC curves were 389 compared using DeLong’s test. 390 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted December 26, 2024. ; https://doi.org/10.1101/2024.12.23.24319587doi: medRxiv preprint