Depression, but not anxiety, is associated with epigenetic age accelerations among Asian older adults

Abstract

Aging is a complex biological process, and biological aging can be quantified by epigenetic clocks. Depression and anxiety are highly prevalent among older adults and are established risk factors for adverse health outcomes. However, their relationships with epigenetic age acceleration (EAA) remain unclear, particularly in Asian populations. Using data from the Diet and Healthy Aging cohort, the present study examined the associations between depressive and anxiety symptoms and EAA within community-dwelling older adults (aged ≥ 60 years, n = 672). Depressive symptoms were assessed using the Geriatric Depression Scale (GDS), and anxiety symptoms were measured using the Geriatric Anxiety Inventory (GAI). Linear mixed-effects models were fitted to all available observations to account for within-person clustering, with additional within-person change analysis conducted among participants with repeated DNA methylation profiles (n = 116). These were followed by rigorous sensitivity analyses to inspect robustness. We found that depressive symptoms, but not anxiety, were robustly associated with higher EAA, primarily indexed by PCPhenoEAA. In fully adjusted models, each standard deviation (SD) increase in depressive symptoms corresponded to a 0.087 SD increase in PCPhenoEAA (β = 0.087, 95% CI [0.023, 0.151], p = 0.008). Participants screening positive for depression (GDS ≥ 5) exhibited, on average, 0.244 SD higher PCPhenoEAA compared with those without depression (β = 0.244, 95% CI [0.027, 0.461], p = 0.030). Despite relatively stable EAA across the follow-up period, within-person change in depressive symptoms was associated with a concomitant increase in PCPhenoEAA. Our findings highlight depression as an important and potentially modifiable factor in delaying biological aging among older Asian adults, highlighting the need for timely screening and interventions to promote healthy aging.

Introduction

Global population aging highlights the importance of healthy aging (HA), which aims to maintain functional ability and well-being in later life. The rising prevalence of depression and anxiety, however, presents major barriers to achieving HA. It is estimated that approximately one in every four older adults globally experiences depression and anxiety [1, 2]. Depression and anxiety are associated with adverse outcomes, including cognitive decline, frailty, vascular disease, and mortality [3–6]. However, the extent to which late-life affective symptom burden relates to biological aging remains insufficiently studied, particularly in Asian populations.

Epigenetic aging

Epigenetic modification is recognized as a primary hallmark of aging [7]. Epigenetic regulation of transcription, gene expression, and downstream nuclear functions involves direct modifications and coordination of DNA, without altering the DNA sequence [8]. A key mechanism here is DNA methylation, which often exhibits tissue-specific and cell-type-specific patterns. Prior literature revealed that advanced chronological age is associated with alterations in these patterns, including de novo hypermethylation at cytosine-phosphate-guanine (CpG) islands and hypomethylation of the global CpGs [8]. Reflecting complex physiological processes, these patterns have directly informed the development of epigenetic clocks that estimate biological-age proxies from information of DNA methylation at CpG sites.

First-generation epigenetic clocks, including HorvathAge (353 CpGs) [9] and HannumAge (71 CpGs) [10], were developed in 2013 to predict chronological age. Subsequent developments have led to second-generation clocks, known as GrimAge (1030 CpGs) and PhenoAge (513 CpGs), which were trained with respective bio-clinical markers to capture mortality and health span phenotypes, as well as third-generation measures, DunedinPace, designed to quantify the rate of aging [11–13]. Overall, epigenetic clocks and related DNA methylation-based biomarkers can be flexibly applied across tissues to index biological aging, with later generations (e.g., second and third generations) showing stronger associations with health-related outcomes than first generations [14]. Epigenetic age acceleration (EAA), defined as the residual from regressing DNA methylation-based age on chronological age, is commonly utilized to indicate whether an individual appears epigenetically older or younger than expected for their chronological age.

Existing studies linking depression and anxiety with epigenetic aging have reported heterogeneous findings across clocks, populations, and symptom measures [15, 16]. Moreover, most evidence comes from Western cohorts, which may limit generalizability to Asian populations whose cultural and ethnic backgrounds differ. To our knowledge, and although associations with psychosocial stress were evaluated [17], no studies to date have examined the late-life anxiety-epigenetic aging association in the older Asian population.

Using data from the Diet and Healthy Aging (DaHA) cohort, the present study evaluated whether depressive and anxiety symptom severity are associated with EAA across multiple epigenetic aging measures in community-dwelling older Asian adults. Furthermore, within-person symptom changes in relation to EAA changes over time were evaluated in the longitudinal analysis. Rigorous sensitivity analyses, including sex-specific trajectories, were performed to strengthen the robustness of the findings.

Methods

Participants were drawn from a prospective cohort study, the Diet and Healthy Aging (DaHA). DaHA was approved by the National University of Singapore Institutional Review Board [reference number 10-517], with written informed consent obtained from all participants. Participants were recruited via door-to-door invitations. Eligible participants were Singaporeans or permanent residents aged 60 years and above who could independently complete dietary records and neurocognitive assessments. Exclusion criteria included terminal illness, dementia, or other conditions affecting cognition. A total of 1010 participants were enrolled between 2011 and 2017, of whom 620 were reassessed between 2017 and 2020, resulting in an average follow-up of 5 years. The details of this cohort are presented elsewhere [18]. The present analysis retained all samples with DNA methylation profiling data.

Sample size and power calculation

Power analyses were performed using the simr and pwr packages in R [19, 20]. Power analysis for analyses using a linear mixed-effects model was estimated via simulation. Assuming a small to medium standardized effect size (β = 0.20) and a two-sided alpha of 0.05, simulation-based analysis suggested that all full models (Model 3) achieved a statistical power of 80% or above with the current sample size. Moreover, assuming a small-to-moderate standardized effect size (f² = 0.25), seven time-varying covariates, and a two-sided α of 0.05, the available sample size (n = 116) provided high statistical power (> 0.95) to detect within-person associations.

Psychological measurements

Depression was assessed using the 15-item Geriatric Depression Scale (GDS-15), with a total score ranging from 0–15; higher scores reflect greater depressive symptomatology. A cutoff of 5 is widely considered for subsyndromal depression [21]. Anxiety was measured using the 20-item Geriatric Anxiety Inventory (GAI-20). The total score ranges from 0–20, where higher scores indicate greater anxiety symptoms. A cutoff of 8 is widely cited to indicate any anxiety disorder [22].

DNA-methylation-based age

Fasting venous blood from consented participants was collected into EDTA-coated tubes by certified phlebotomists. Bisulfite conversion was performed to convert unmethylated cytosines to uracils for subsequent amplification and methylation profiling using the Zymo EZ DNA methylation Kit (Zymo, Irvine, CA). Genome-wide DNA methylation was evaluated using Infinium Methylation EPIC v1.0 Beadchip (850 K array, Illumina Inc., San Diego, CA, USA).

The IDAT files were first imported into RStudio using the minfi package. Data preprocessing was conducted with Enmix, Sesame, and DMRcate packages. Initial quality check indicated sound bisulfite conversion (Appendix Figure A1 & A2) and single-base extension (Appendix Figure A3). An out-of-band (OOB) background correction was first applied, followed by dye-bias and probe bias corrections. Intensity distributions across samples were normalized using the quantile method (Q1). This pipeline aligned with the published benchmarks, revealing optimized intraclass correlation across diverse epigenetic clocks and surrogates. The P-value matrix was derived using an out-of-band after-hybridization (pooBAH) method, which is effective at controlling germline deletion-related artifacts [23]. Single-nucleotide polymorphism (SNP) and cross-hybridization were handled using the “rmSNPandCH” function from the DMRcate package, with a 5-base-pair gap and a minor allele frequency of 5%. As quality control, samples with detected p-values above 0.05 in more than 10% of their probes (n = 3) and a mismatched sex were excluded from downstream analysis. Moreover, probes with p-values above 0.05 in more than 5% of samples were deemed low-quality and removed. Missing values were replaced with mean GSE40279 beta values for epigenetic age estimation, where necessary.

A total of six clocks, covering first to third generations, were calculated and evaluated [9–13, 24]. To mitigate technical noise and enhance reliability, principal component-based ages were calculated for HorvathAge, Horvath2Age, HannumAge, PhenoAge, and GrimAge per protocol [25]. The mean values for five technical duplicates were calculated and retained for downstream analysis. Epigenetic Age Acceleration (EAA), the residual from regressing epigenetic age on chronological age, was subsequently derived using the lm() function. Individuals with positive EAA exhibit an accelerated aging trajectory relative to their chronological age within the cohort, whereas samples with negative EAA suggest comparatively decelerated aging. DunedinPACE is interpreted as the aging pace relative to the calendar year.

Cell-composition analysis

Using the pre-processed β matrix, twelve immune cell types were deconvolved using the cent12CT.m reference from the EpiDISH package. These included naïve and mature B cells, naïve and memory CD4⁺ T cells, naïve and memory CD8⁺ T cells, T-regulatory cells, NK-cells, Neutrophils, Monocytes, Eosinophils, and Basophils. To mitigate multicollinearity, cell type proportions were further subjected to principal component analysis. The top four components, explaining 75% of total variance, were adjusted in the sensitivity analysis.

Covariates

A comprehensive list of covariates was adjusted to isolate the association between depression/anxiety and epigenetic age acceleration. Covariate selection was informed by prior literature targeting late-life depression [15, 16, 26] and data availability. Sex was coded as binary (0 = males, 1 = females). Educational attainment was adjusted using total years of schooling. Ethnicity was adjusted as binary (0 = Chinese, 1 = non-Chinese), reflecting the predominance of the Chinese population in the cohort. Marital status was self-reported, with married coded as the reference. Housing types were self-reported and recoded into three categories: lower tier (1–3 room Housing and Development Board (HDB)), middle tier (4–5 room HDB), and upper tier (executive maisonette, condo, landed).

Body mass index (BMI) was calculated from measured height and weight. A composite score of chronic conditions was derived from self-reported diagnoses. Frequencies of physical, social activity, smoking, and alcohol consumption were self-reported using an ordinal scale (1 = Never or rarely, 2 = More than once a month but less than once a week, 3 = one to three times a week, 4 = four to six times a week, 5 = daily). Due to limited variability, alcohol consumption and smoking were recoded into binary (0 = Never or rarely, 1 = at least monthly). No participants reported the use of antidepressants or benzodiazepines. Chronological age was not adjusted as it is orthogonal to EAA.

Statistical analysis

All statistical analyses were performed in RStudio with a predetermined seed. A two-sided alpha of 0.05 was applied for hypothesis testing. Descriptive statistics were stratified by sex, with continuous variables summarized as mean ± standard deviation (SD) and categorical variables as frequencies and percentages (n [%]). Missingness in covariates was assumed to be missing at random (MAR) and imputed using multivariate imputation by chained equations (MICE; predictive mean matching; five iterations). No imputation was performed for exposure or outcome variables. Prior to modeling, correlation matrices were examined to assess relationships among variables.

To evaluate associations between depression/anxiety severity and EAAs, linear mixed-effects models were fitted to all available observations (n = 672). Random intercepts were included to account for within-cluster correlation among participants, who contributed to the data twice (n = 116). The significance of the random effect was assessed using likelihood ratio tests comparing models with and without random intercepts. Time was initially modeled as a fixed effect to assess linear change across waves and was removed should no linear growth be observed.

Models were fitted incrementally. Model 1 was unadjusted; Model 2 was adjusted for demographic and health-related factors; and Model 3 (full model) was adjusted for additional lifestyle covariates. To reinforce robustness, parametric resampling (R = 10,000) was applied to obtain the bootstrapped CI using the percentile method. Symptom severity was additionally analyzed using clinically meaningful cutoffs to facilitate interpretation (GDS ≥ 5; GAI ≥ 8). Where appropriate, diagnostic plots were provided in the supplementary materials to ensure that key assumptions of linear regression were satisfied.

Longitudinal within-person change analyses were further performed among participants with repeated measurements at both waves (n = 116; 232 observations) to examine whether changes in symptom severity were associated with concomitant changes in EAA over the follow-up period. The change scores were calculated as follow-up minus baseline values for epigenetic age acceleration (Δ_EAA) and depressive or anxiety symptoms (Δ_Dep, Δ_Anx) for respective participants. Linear regression models were subsequently fitted with Δ_EAA as the outcome and Δ_symptoms as predictors. Models were adjusted for time-varying covariates, including changes in BMI, composite number of chronic diseases, and the frequency of lifestyle activity engagement. As a sensitivity analysis, models were further adjusted for changes in cell compositions.

Drawing on existing literature [15, 27] and its biological relevance [14], depressive symptoms and a second-generation epigenetic clock, PhenoAge, were designated as the primary exposure and outcome, respectively. Analyses involving other epigenetic clocks and anxiety-related associations were deemed exploratory. False Discovery Rate (FDR) correction using the Benjamini-Hochberg (BH) method was applied across all secondary analyses within the fully adjusted models to mitigate multiple testing.

Sensitivity analysis

A series of sensitivity analyses was performed to assess the robustness of the findings, including additional adjustment for blood cell composition, evaluation of sex-specific effects, robust regression to inspect influential observations, and inspection of potential nonlinear associations using spline models. Detailed methodologies and results are presented in the supplementary materials.

Results

The analytic sample comprised 672 older Asian adults from the DaHA cohort, with a mean age of 70 years (± 5.8). Most participants were married (68.2%) and resided in 4–5 room HDB flats (72.9%). Over one-third of participants had no formal education (35.6%). Both depressive and anxiety were low: the mean GDS-15 was 1.30 (± 2.0) and the mean GAI-20 score was also 1.30 (± 2.7). A total of 56 participants (8.3%) screened positive for depression (GDS ≥ 5), with 28 participants screening positive for anxiety (GAI ≥ 8). Most participants reported not drinking or smoking or did so rarely, and engaged in regular physical and social activities. Epigenetic age acceleration across clocks was centered around zero, with standard deviations ranging from 2.7–5.4. Participants’ characteristics are presented in Table 1. As shown in Fig. 1, chronological age was strongly and positively correlated with all epigenetic age measures and most surrogate markers underlying PCGrimAge (all p < 0.001), with ρ ranging from 0.159 (DunedinPACE) to 0.940 (PCTIMP1). Depressive symptoms were weakly and negatively correlated with both PCHorvathAge (ρ = −0.082, p = 0.03) and PCGrimAge (ρ = −0.086, p = 0.027), whereas no consistent correlation was observed for anxiety symptoms. Estimated correlation coefficients are available in the supplementary material.

Table 1.

Characteristics of analytic sample.

	male (N = 188)	female (N = 484)	Overall (N = 672)
Chronological age (years)
Mean (SD)	71 ( ± 6.0)	70 ( ± 5.7)	70 ( ± 5.8)
Total years in school
Mean (SD)	8.3 ( ± 4.1)	5.8 ( ± 4.1)	6.5 ( ± 4.2)
Highest education attainment n(%)
No formal education	41 (21.8%)	198 (40.9%)	239 (35.6%)
Primary	66 (35.1%)	153 (31.6%)	219 (32.6%)
Secondary	45 (23.9%)	106 (21.9%)	151 (22.5%)
Pre Univ or Polytechnic	22 (11.7%)	20 (4.1%)	42 (6.3%)
University	14 (7.4%)	7 (1.4%)	21 (3.1%)
Marital status n(%)
Married	159 (84.6%)	299 (61.8%)	458 (68.2%)
Single	9 (4.8%)	12 (2.5%)	21 (3.1%)
Divorced/Seperated	6 (3.2%)	33 (6.8%)	39 (5.8%)
Widowed	14 (7.4%)	140 (28.9%)	154 (22.9%)
Housing tier
middle tier	135 (71.8%)	355 (73.3%)	490 (72.9%)
lower tier	27 (14.4%)	77 (15.9%)	104 (15.5%)
upper tier	26 (13.8%)	52 (10.7%)	78 (11.6%)
Depressive symptoms (GDS-15)
Mean (SD)	1.2 ( ± 1.8)	1.3 ( ± 2.1)	1.3 ( ± 2.0)
Depression status (GDS ≥ 5)
Normal	173 (92.0%)	443 (91.5%)	616 (91.7%)
Positive screen	15 (8.0%)	41 (8.5%)	56 (8.3%)
Anxiety symptoms (GAI-20)
Mean (SD)	0.94 ( ± 2.4)	1.4 ( ± 2.8)	1.3 ( ± 2.7)
Anxiety statis (GAI ≥ 8)
Normal	183 (97.3%)	461 (95.2%)	644 (95.8%)
Positive screen	5 (2.7%)	23 (4.8%)	28 (4.2%)
BMI
Mean (SD)	24 ( ± 3.4)	24 ( ± 4.1)	24 ( ± 3.9)
Number of chronic Diseases
Mean (SD)	3.0 ( ± 2.2)	2.6 ( ± 1.7)	2.7 ( ± 1.9)
Physical activity frequency n(%)
Never or rarely	28 (14.9%)	54 (11.2%)	82 (12.2%)
> 1 in one month but < 1 in a week	1 (0.5%)	15 (3.1%)	16 (2.4%)
1–3 times a week	34 (18.1%)	114 (23.6%)	148 (22.0%)
4–6 times a week	39 (20.7%)	134 (27.7%)	173 (25.7%)
daily	86 (45.7%)	167 (34.5%)	253 (37.6%)
Social activity frequency n(%)
Never or rarely	45 (23.9%)	48 (9.9%)	93 (13.8%)
> 1 in one month but < 1 in a week	27 (14.4%)	65 (13.4%)	92 (13.7%)
1–3 times a week	66 (35.1%)	224 (46.3%)	290 (43.2%)
4–6 times a week	26 (13.8%)	79 (16.3%)	105 (15.6%)
daily	24 (12.8%)	68 (14.0%)	92 (13.7%)
Smoking status n(%)
Never or rarely	170 (90.4%)	476 (98.3%)	646 (96.1%)
Regular	18 (9.6%)	8 (1.7%)	26 (3.9%)
Alcohol eonsumption n(%)
Never or rarely	156 (83.0%)	446 (92.1%)	602 (89.6%)
Regular	32 (17.0%)	38 (7.9%)	70 (10.4%)
PCHorvathEAA (years)
Mean (SD)	1.2 ( ± 5.1)	−0.48 ( ± 4.7)	0.00 ( ± 4.9)
PCHorvath2EAA (years)
Mean (SD)	0.79 ( ± 5.7)	−0.31 ( ± 5.3)	0.00 ( ± 5.4)
PCHannumEAA (years)
Mean (SD)	1.7 ( ± 4.8)	−0.68 ( ± 4.5)	0.00 ( ± 4.7)
PCPhenoEAA (years)
Mean (SD)	1.4 ( ± 5.6)	−0.53 ( ± 5.1)	0.00 ( ± 5.3)
PCGrimEAA (years)
Mean (SD)	1.9 ( ± 2.8)	−0.73 ( ± 2.4)	0.00 ( ± 2.7)
DunedinPACE
Mean (SD)	1.0 ( ± 0.11)	1.0 ( ± 0.12)	1.0 ( ± 0.12)

Fig. 1. Correlation structure among epigenetic aging measures, GrimAge-related DNA methylation–based surrogate markers, and key covariates.

Heatmap displaying Spearman’s rank correlation coefficients (ρ) between epigenetic clocks (PC-Horvath, PC-Horvath2, PC-Hannum, PC-PhenoAge, PC-GrimAge, DunedinPACE), GrimAge-related DNA methylation-based surrogate markers (e.g., DNAmPACKYRS, DNAmADM, DNAmB2M, DNAmCystatin C, DNAmGDF15, DNAmLeptin, DNAmPAI1, DNAmTIMP1), and covariates (age, years of schooling, BMI, number of chronic conditions, depressive symptoms [GDS], anxiety symptoms [GAI], and cell-composition principal components PC1–PC7). Color reflects the magnitude and direction of correlations (blue = positive, red = negative). Asterisks denote statistical significance (* p < 0.05; ** p < 0.01; *** p < 0.001).

The random-effect analysis demonstrated high intra-cluster correlations (all ICCs > 0.70), indicating relatively stable epigenetic age trajectories among individuals with repeated measures and substantial between-individual differences. Full estimated parameters are presented in supplemental tables (S1 & S2). Bootstrapped estimates for primary models are presented in histograms (see Supplemental figures S4 & S5). These estimates were generally distributed around the original point estimates.

Among the clocks evaluated, the most pronounced and significant association was observed for PCPhenoEAA, with its effect sizes increasing incrementally with progressive covariates adjustment (Fig. 2a). In the full model, each standard deviation (SD) increase in depressive score corresponded to an additional 0.087 SD of PCPhenoEAA (Model 3: β = 0.087, 95% CI [0.023, 0.151], p = 0.008). Using a clinically meaningful cutoff (GDS ≥ 5), participants screening positive for depression exhibited, on average, 0.244 SD higher PCPhenoEAA relative to those without depression (β = 0.244, 95% CI [0.027, 0.461], p = 0.030). A similar pattern was also observed for DunedinPACE, indicating an accelerated aging pace among individuals with a positive screen for depression (β = 0.244, 95% CI [0.018, 0.469], p = 0.036); however, this association was not significant after correction for multiple testing. No consistent associations were identified between depression severity and the remaining clocks (all p > 0.05), although estimates were mostly directionally positive (β range: 0.04–0.14).

Fig. 2. Forest plot summaries of linear mixed-effects models examining associations between depression, anxiety severity and epigenetic age acceleration (EAA) (n = 672 observations).

a Associations between depressive symptoms and EAA across epigenetic clocks. b Associations between anxiety symptoms and EAA across epigenetic clocks. Principal component-based ages were calculated for first and second generations. Results were rounded into three decimals. Points represent standardized regression coefficients (i.e., the estimated change in EAA, in SD units, per 1-SD increase in depression or anxiety, or for comparisons based on clinical cutoffs (e.g., GDS ≥ 5; GAI ≥ 8). Error bars represent 95% confidence intervals (CIs). An asterisk (*) indicates statistical significance (p < 0.05). For Model 3 and Model 6 (fully adjusted models), bootstrapped CIs are shown. p values for the fully adjusted models were corrected for multiple testing using the Benjamini–Hochberg procedure (see Methods). The vertical dashed line denotes the null. Model 1: Crude model; Model 2: Model adjusted for demographic and health covariates (sex, education, marital status, ethnicity, housing type, number of chronic diseases, BMI) Model 3: Model 2 further adjusted for lifestyle covariates (physical activity, social activity, alcohol consumption, smoking status) Model 4: Crude model with symptoms categorized into clinical cutoff (GDS ≥ 5; GAI ≥ 8). Model 5: Model 4 adjusted for demographic and health-related covariates; Model 6: Fully adjusted model with symptoms operationalized using the clinical cutoff.

Conversely, anxiety severity was not significantly associated with epigenetic age acceleration across any clocks in the analytic sample. Despite the non-significance, the continuous score demonstrated positive trends within most clocks (Model 3, β range: 0.02–0.04) (Fig. 2b). The estimated trends for participants screening positive for anxiety (GAI ≥ 8) were instead directionally negative across clocks (Model 6, β range: −0.54–0.03), whereas neither approached statistical significance.

Few covariates demonstrated significant associations with EAA. Biological sex appeared to be a consistent predictor, with females, on average, exhibiting lower EAA than males. As these covariate effects were not the primary focus, detailed results are provided in the supplementary material.

Longitudinal analysis

The longitudinal analysis was restricted to participants with DNA methylation profiles at both waves (n = 116; 232 observations), resulting in a median follow-up duration of 4.76 years (IQR: 1.19). Results are presented in supplemental tables (S5-S8). Both depressive (β = −0.28, 95% CI [−0.50, −0.05]) and anxiety symptoms (β = −0.23, 95% CI [−0.44, −0.03]) decreased at follow-up relative to baseline, indicating overall remission in the sample. Although high ICCs in mixed models indicated generally stable EAA trajectories within individuals, the model revealed that within-person changes in depressive symptom severity tracked concomitant fluctuations in PCPhenoEAA over time. Specifically, each SD increase in depressive symptoms was associated with a 0.19 SD increase in PCPhenoEAA across time (β = 0.19, 95% CI [0.01, 0.36], p = 0.043), after adjusting for time-varying covariates. Additionally, this association remained robust in a sensitivity analysis, where within-person changes in cell composition (Δ_Cell-pcs) were further adjusted (β = 0.18, 95% CI [0.01, 0.36], p = 0.040).

Exploratory analyses suggested that within-person increases in depressive symptoms were also associated with changes in PCHorvath2EAA and PCHannumEAA after adjusting for cell-composition PCs. However, these associations were not significant after multiple testing correction. Regarding anxiety, no within-person associations were observed for second or third-generation epigenetic clocks. While some first-generation clocks showed nominal associations, none withstood multiple testing correction and were therefore considered hypothesis-generating.

Sensitivity analysis

Results were integrated into the supplementary material. Post hoc analyses were first performed to inspect the robustness of the depression-PCPhenoEAA association. The assumption of linearity and normality was satisfied in the model, with no evidence of heteroscedasticity (supplemental figure S9). Robust regression using MM estimators was subsequently refitted to ensure observed patterns were not confounded by potential outliers. Most observations received full or near-full weight (n = 647 [96.3%]), with only a small proportion down-weighted. The estimates were largely comparable to the original analyses, regardless of whether depression was modeled continuously or operationalized using clinically cutoffs. The original estimates remained significant and directionally consistent after further adjustment for cell-composition principal components (β = 0.06, p = 0.001) and cognitive status (β = 0.07, p = 0.009).

In evaluating effect modification by biological sex, borderline significant interactions were observed for both PCPhenoEAA and PCGrimEAA. Likelihood ratio tests comparing these models without an interaction term suggested slight model improvements, although none reached significance. To assess the robustness of these findings, parametric bootstraps (10,000 iterations) were conducted. The resulting 95% percentile-based confidence intervals for interaction term excluded the null (PCPhenoEAA: 95% CI [0.011, 0.222], PCGrimEAA: 95% CI [0.017, 0.212]). These collectively indicated potential sex-specific patterns in depression-related epigenetic aging trajectories. Conversely, no consistent sex interactions were observed for the remaining clocks and anxiety-related analyses.

Natural cubic splines adjusted for covariates did not outperform the original linear models (χ² (2) = 0.0854–3.239, all p > 0.05) when estimating marginal effects. As shown in supplemental figures S7 and S8, the predicted splines closely resembled straight lines across EAAs. Of note, the relationship between depressive and anxiety symptoms and EAA was rather monotonic in the general older population.

Discussion

Summary of findings

This study, to our knowledge, is the first to integratively examine depressive and anxiety symptoms in relation to epigenetic age acceleration (EAA) within a community sample of older Asian adults. Results demonstrated that greater depressive symptoms were consistently associated with accelerated epigenetic aging, particularly as indexed by PCPhenoEAA, while anxiety symptoms showed no consistent association across all epigenetic clocks. These findings were observed across covariate-adjusted models and remained stable across extensive sensitivity analyses, including robust regression and additional adjustment for methylation-derived cell types.

In addition to cross-sectional associations, the within-person change analyses in participants with repeat methylation profiling afforded additional insights. Although EAA demonstrated relatively high within-person stability over the follow-up period, increases in depressive symptoms tracked concomitant increases in PCPhenoEAA over time, even after adjusting for time-varying covariates. Of note, this indicates that the depression-EAA association cannot be solely attributed to between-person differences.

The observed pattern for depressive symptoms aligns with evidence from Western cohorts linking depression-related phenotypes to EAA [18]. In contrast, findings related to anxiety were inconclusive. This null finding may be due to limited variability in anxiety symptoms within the sample. When operationalized using clinical cutoffs (GAI ≥ 8), estimated trends across clocks were directionally negative; however, none reached significance. This direction aligns with prior research, which reports suppressed EAA among individuals with diagnosed anxiety disorders in younger adult cohorts [28]. Differences in age range and phenotype may partially explain the lack of replication in the present study.

Exploratory analyses revealed potential sex-dependent trajectories for both PCPhenoEAA and PCGrimEAA in the context of depression. Interaction tests were borderline significant, and bootstrap confidence intervals that excluded the null. Model-based marginal predictions were consistent with a steeper depression-EAA pattern among females over males. These interaction findings, however, were not consistent across others and were not robust when depression was modeled using clinical screening cutoffs. Given the multiple testing, we interpreted these as hypothesis-generating.

Potential mechanism (hypothesis-generating)

Although biological mechanisms were not explicitly evaluated in the present study, the observed depression-EAA patterns are broadly consistent with prior literature linking depressive phenotypes to dysregulation of stress, immune, and metabolic symptoms [29–35]. Particularly, methylation-derived C-reactive protein (CRP), which features prolonged inflammatory response to exposure, has been shown to associate with more widespread brain structural alterations and depressive symptom dimensions, supporting the relevance of epigenetic inflammation biomarkers for downstream aging phenotypes [36].

The schematic summary (Fig. 3) is provided here to summarize several candidate pathways, proposed in prior experimental studies, including dysregulation of stress-responsive circuits (e.g., HPA-axis signaling) [3a], altered D2/D3 receptor expression and impaired reward sensing [3b], and cumulative allostatic load [3c] [34]. These biologically plausible mechanisms may contribute to epigenetic signatures captured by second-generation clocks, such as PhenoAge, which were trained on morbidity and mortality.

Fig. 3. Schematic summary of potential biological mechanisms explaining depression-EAA association.

a HPA axis dysregulation: Depression can dysregulate HPA axis and increases serum glucocorticoids concentrations. Excessive activation of MRs on SMCs induces vascular stiffness, fibrosis, and inflammation. Binding of GRs initially upregulates anti-inflammatory factors and downregulate pro-inflammatory cytokines through NF-κB inhibition. However, depression has been found to be associated with hypermethylation at NR3C1 promoter, attenuating the overall GRs expression, which induces chronic inflammation (inflammaging). b Dopaminergic dysfunction. Depression is associated with enhanced expression of D2/D3 receptors, which suggests downregulation of the mesolimbic and nigrostriatal pathways. These result in reduced reward sensing and motivation of achieving goals, impaired motor and cognitive functions. c Allostatic load: An accumulation of physiological wear and tear results from exogenous stress such as depression. Studies suggest positive association between depressive symptoms and allostatic load in community-dwelling older adults, which in turn is associated with increased frailty, chronic inflammation, and cognitive decline. d CRF-1 coupling mechanism: CRF is hypersecreted in the context of depression. One target of CRF is CRF-1 receptors located at locus coeruleus. Higher proportion of binding of CRF to CRF-1 in females are coupled to G-protein coupled receptors, which activates cAMP PKA pathway that phosphorylates target genes, further increasing CRF sensitivity. Greater proportion of binding of CRF to CRF-1 in males are coupled to beta-arrestins that internalize CRF-1 receptors to increase CRF resilience under CRF hypersecretion. e DNMT expression differences: Expression of Dnmt1 was found to be lower in MDD patients regardless of sex, suggesting inability to maintain original methylation patterns. While Dnmt3B expression was increased in MDD patients, such extents were pronounced in females MDD, suggesting uneven de-novo methylation that altering transcription and function. HPA, Hypothalamic-pituitary-adrenal; SMC, smooth muscle cell; GRs, glucocorticoids receptors; MRs, mineralocorticoid receptors; CRF: corticotrophin releasing factor; cAMP-PKA: cyclic adenosine monophosphate-protein kinase A; DNMT: DNA methyltransferases. Figure generated in Biorender.

In parallel, potential sex-related biological heterogeneity in the corticotropin-releasing hormone (CRH) receptor coupling mechanism [37] and the reported methylation level in post-mortem brain tissue [37] are also highlighted (Figs. 3d, 3e), and may facilitate our contextualization of the exploratory sex-by-depression interactions observed here. However, given the modest effect sizes and absence of direct measures of endocrine or molecular signatures, these patterns are interpreted as hypothesis-generating that warrant extension in studies explicitly designed to test sex-specific biological mediators.

Limitations

This study possesses several strengths, including the use of principal component-based epigenetic clocks to reduce noise, modelling of repeated observations, and a comprehensive set of robustness checks. However, several limitations are acknowledged. First, observational design limits causal inference, and residual confounding, such as unadjusted inflammatory markers, sleep hygiene, and psychosocial stressors and resilience, cannot be excluded. Second, the longitudinal analysis was limited to participants with DNA methylation data available at both waves. Although the specific reasons for missingness were not systematically collected, many reported having routine medical screenings with general practitioners, and considered repeat blood draws unnecessary. Hence, the estimates may be sensitive to attrition and selection processes. Third, symptom distributions were generally right-skewed in this non-clinical cohort, which may reduce power to detect anxiety associations, as the cohort was relatively healthy. Lastly, the ability to explore the effects of clinical cases versus non-clinical cases on aging acceleration is limited in this non-clinical cohort.

Conclusion

In this cohort of community-dwelling older Asian adults, late-life depression was associated with epigenetic age acceleration, most consistently for the second-generation clock, PCPhenoEAA, whereas anxiety symptoms were not robustly associated with any epigenetic age acceleration. Whether alleviating depressive symptoms alters molecular aging trajectories requires additional longitudinal studies and/or intervention studies to scaffold causality.

Author contributions

LF was the principal investigator of the study. JG and LF conceived and designed the study. JG performed the coding and data analysis, and drafted the first manuscript. KXY contributed to data wrangling, and item cross-checking. All other authors provided invaluable feedback and comments that helped refine the manuscript.

Funding

This study was supported by the National University of Singapore Virtual Institute for the Study of Aging [VG-8]; the Alice Lim Memorial Fund, Singapore [ALMFA/2010]; and the National Medical Research Council of Singapore ([NMRC/TA/0053/2016], [NMRC/CSA/INV/0009/2022]) awarded to Dr Feng Lei.

Data availability

The data analyzed in this study were obtained from the Diet and Healthy Aging study and are not yet publicly available due to ethical and privacy constraints. However, the data and source code may be made available upon reasonable request or through collaborative projects with the corresponding author (fenglei2020sg@nus.edu.sg), subject to appropriate approvals.

Code availability

The code used for analysis and figure generation is available from the corresponding author upon reasonable request, subject to ethical approval.

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

All methods in the present study were conducted in accordance with relevant guidelines and regulations. This study was approved by the Institutional Review Board of National University of Singapore [reference number 10-517]. Written informed consent was obtained from all participants.

Supplementary information

The online version contains supplementary material available at 10.1038/s41380-026-03588-2.