1. Cell Biology
  2. Genetics and Genomics
Download icon

An integrative study of five biological clocks in somatic and mental health

  1. Rick Jansen  Is a corresponding author
  2. Laura KM Han
  3. Josine E Verhoeven
  4. Karolina A Aberg
  5. Edwin CGJ van den Oord
  6. Yuri Milaneschi
  7. Brenda WJH Penninx
  1. Department of Psychiatry, Amsterdam UMC, Vrije Universiteit Amsterdam, Amsterdam Public Health Research Institute and Amsterdam Neuroscience, Netherlands
  2. Center for Biomarker Research and Precision Medicine, Virginia Commonwealth University, United States
Research Article
  • Cited 1
  • Views 1,340
  • Annotations
Cite this article as: eLife 2021;10:e59479 doi: 10.7554/eLife.59479

Abstract

Biological clocks have been developed at different molecular levels and were found to be more advanced in the presence of somatic illness and mental disorders. However, it is unclear whether different biological clocks reflect similar aging processes and determinants. In ~3000 subjects, we examined whether five biological clocks (telomere length, epigenetic, transcriptomic, proteomic, and metabolomic clocks) were interrelated and associated to somatic and mental health determinants. Correlations between biological aging indicators were small (all r < 0.2), indicating little overlap. The most consistent associations of advanced biological aging were found for male sex, higher body mass index (BMI), metabolic syndrome, smoking, and depression. As compared to the individual clocks, a composite index of all five clocks showed most pronounced associations with health determinants. The large effect sizes of the composite index and the low correlation between biological aging indicators suggest that one’s biological age is best reflected by combining aging measures from multiple cellular levels.

Introduction

Aging can be conceptualized in different ways. While chronological age is measured by date of birth, biological age reflects the relative aging of an individual’s physiological condition. Biological aging can be estimated by various cellular indices (López-Otín et al., 2013). Commonly used indices are based on telomere length, DNA methylation patterns (epigenetic age), variation in transcription (transcriptomic age) as well as alterations in the metabolome (metabolomic age) and in the proteome (proteomic age) (see Han et al., 2019Xia et al., 2017 and Jylhävä et al., 2017 for recent reviews). Biological aging is defined as the residuals of regressing predicted biological age on chronological age: a positive value indicates that the biological age is larger than the chronological age. Advanced biological aging (i.e. an increased biological clock) has been associated to poor somatic health, including the onset of aging-related somatic diseases such as cardiovascular disease, diabetes, and cognitive decline (Xia et al., 2017). Advanced biological aging has also been correlated to mental health: childhood trauma (Li et al., 2017), psychological stress, and psychiatric disorders (Darrow et al., 2016; Han et al., 2018). Specifically, telomere length has been most extensively researched and was found to be shorter in various somatic conditions (Jin et al., 2018), all-cause mortality (Mons et al., 2017; Wang et al., 2018) and a range of psychiatric disorders (Lindqvist et al., 2015). Advanced epigenetic aging has also been linked to worse somatic health, mortality (Marioni et al., 2015), depressive disorder (Han et al., 2018; Whalley, 2017), and post-traumatic stress disorder (Wolf et al., 2018), although some studies have found associations with the opposite direction of effect (Verhoeven et al., 2018; Boks et al., 2015). Advanced transcriptomic aging was found in those with higher blood pressure, cholesterol levels, fasting glucose, and body mass index (BMI) (Peters et al., 2015). Advanced metabolomic aging increases risk on future cardiovascular disease, mortality, and functionality (Akker et al., 2019).

While all biological clocks aim to measure the biological aging process, there is limited evidence for cross-correlations among different clocks. Belsky and colleagues (Belsky et al., 2017) recently showed low agreement between eleven quantifications of biological aging including telomere length, epigenetic aging, and biomarker-composites. In contrast, Hastings et al., 2019 showed relatively strong correlations (r > 0.50) between three physiological composite biological clocks (i.e. homeostatic dysregulation, Klemer and Doubal’s method and Levine’s method), but not with telomere length. Other studies showed that telomere length was not correlated with epigenetic aging (Han et al., 2018; Marioni et al., 2018), although cell type composition adjustments revealed a modest association (Chen et al., 2017). Further, both Hannum and Horvath epigenetic clocks (Hannum et al., 2013; Horvath et al., 2012) showed modest correlations to a transcriptomic clock.

Most previous studies, however, have separately considered the relation between a single biological clock and different somatic and mental health conditions. To date, extensive integrated analyses across multiple cellular and molecular aging markers in one study are lacking and it remains unknown to what extent different biological clocks are similarly associated to different health determinants. In addition, most studies did not examine health in its full range and, consequently, whether both somatic and mental health are associated with biological aging remains elusive. As it is unlikely that a single biological clock can fully capture the complexity of the aging process (Cole et al., 2019), a composite index, that integrates the different biological clocks and thereby aging at several molecular levels, may reveal the strongest health impact. Therefore, there is an additional need to integrate different biological clocks and test whether such a ‘composite clock’ outperforms single biological blocks in its association with health determinants.

To develop a better understanding of the mechanisms underlying biological aging, this study aimed to examine (1) the intercorrelations between biological aging indicators based on different molecular levels ranging from DNA to metabolites, namely telomere length, epigenetic, transcriptomic, proteomic and metabolomic clocks; (2) the relationships between different biological aging indicators with both somatic and mental health determinants; and (3) whether a composite biological clock outperforms single biological aging indicators in its association with health. For the five biological aging indicators and the composite clock, associations were computed with a wide panel of lifestyle (e.g. alcohol use, physical activity, smoking), somatic health (functional indicators, BMI, metabolic syndrome, chronic diseases) and mental health (childhood trauma, depression status) determinants.

Results

Sample characteristics

To create indicators for biological aging we used whole blood derived measurements from the Netherlands Study of Depression and Anxiety (NESDA) baseline assessment: telomere length (N = 2936), epigenetics (DNA methylation, N = 1130, MBD-seq, 28M CpGs), gene expression (N = 1990, Affymetrix U219 micro arrays, >20K genes), proteomics (N = 1837, Myriad RBM DiscoveryMAP 250+, 171 proteins) and metabolites (N = 2910, Nightingale Health platform, 231 metabolites), with 653 overlapping samples (see Table 1 for sample characteristics). Each subsample included around 66% female, with mean age of around 42 years.

Table 1
Sample description.
Telomere
Length
Epigenetic
Aging
Transcriptomic
Aging
Proteomic
Aging
Metabolomic
Aging
Composite
Index
# Subjects29361130199018372910653
DemographicSex (%female)66.0065.0067.0067.0066.0066.00
Education years (mean)12.1511.9312.0712.0712.1511.71
Age (mean)41.8141.5338.7141.3741.9441.23
LifestyleAlcohol use (units per week, mean)6.246.546.386.396.296.48
Smoking (pack years, mean)11.0011.4311.8410.3711.1210.90
Physical activity (MET minutes per week, mean)3679.723638.543729.203741.003668.133525.05
Somatic HealthBMI (mean)25.6025.6725.6825.6625.6025.82
Physical disability (score, mean)24.4029.4526.0023.2224.4530.27
Lung capacity (PEF in liter/minute, mean)477.74479.75478.42477.19477.23475.23
Hand grip strength (kg, mean)37.0637.7737.0837.4637.0537.74
Cardiometabolic disease (%cases)181818181817
Respiratory disease (%cases)9999910
Musculoskeletal disease (%cases)1010109109
Digestive disease (%cases)999898
Neurological disease (%cases)323332
Endocrine disease (%cases)333334
Cancer (%cases)787778
Metabolic syndrome (# components, mean)1.361.391.371.331.361.41
# Chronic diseases (mean)0.610.620.620.580.610.63
Mental HealthCurrent MDD (%cases)277234262776
Depression severity (IDS, mean)21.4625.8022.9120.9621.4826.67
Childhood Trauma (score from 0-4, mean)0.910.971.000.870.921.01

Computing biological clocks

The methods for creating the biological clocks are described in detail in the Materials and methods section. In brief, for each of the four omics measures (epigenetic, transcriptomic, metabolomic and proteomic) we estimated biological age using ridge regression and cross validation (see Figure 1 for study design). As telomere length values usually decline with increasing chronological age, this indicator was multiplied by −1 to be able to compare directions of effects consistent with the other biological clocks. Correlations between chronological age and predicted biological age were 0.30 for telomere length, 0.95 for epigenetic age, 0.72 for transcriptomic age, 0.85 for proteomic age, and 0.70 for metabolomic age (Figure 1). For each omics-based biological clock, biological aging is defined as the residuals of regressing predicted biological age on chronological age: a positive value means that the biological age is larger than the chronological age. The individual clocks residualized for chronological age are also referred to as biological aging indicators. Correlations between biological aging indicators, corrected for sex, are presented in Figure 2. Correlations were significant for 3 out of 10 pairs; proteomic vs metabolomic aging (r = 0.19, p=2e-16), transcriptomic vs epigenetic aging (r = 0.15, p=3e-06) and transcriptomic vs proteomic aging (r = 0.08, p=2e-06).

Study design.

The upper part of the figure shows the five biological layers. From each of the four omics layers (epigenetic, transcriptomic, proteomic, and metabolomic data), biological age was estimated, and biological age was regressed on age to obtain measures of biological aging. Only telomere length was not age-regressed. The five biological aging indicators were associated with multiple demographic, lifestyle, somatic health and mental health determinants.

Correlations between the biological aging indicators.

The heatmap represents Spearman rank correlations between the five biological aging indicators, all corrected for sex. Out of 10 pairs, three are significant: transcriptomic vs epigenetic aging, metabolomic vs proteomic aging and proteomic vs transcriptomic aging. All biological aging indicators were age-regressed, only telomere length was not.

Associations between individual biological aging indicators and health determinants

For each of the five biological aging indicators, we computed associations with several demographic (sex, education), lifestyle (physical activity, smoking, alcohol use), somatic health (BMI, hand grip strength, lung function, physical disability, chronic diseases), and mental health (current depression, depression severity, childhood trauma) determinants. Except for proteomic aging, sex was associated with all biological aging indicators: women were biologically younger than men (p=3e-4 for telomere length, p=5e-4 for epigenetic aging, p=4e-11 for transcriptomic aging, p=1e-5 for metabolomic aging). Education was not associated with any biological aging indicator. We controlled for sex by using it as a covariate in all following models (except for in the model where sex was the outcome). Table 2 and Figure 3 give an overview of all associations. Correction for multiple testing was done using permutation-based FDR (Materials and methods), resulting in a p-value threshold of 2e-2 for an FDR of 5% for all tests.

Forest plot of associations between biological aging and health determinants.

For each of the associations between biological aging indicators and health determinants, the standardized beta and standard deviation derived from linear models were plotted. The significant associations (p<2e-2, FDR < 5%) are shown with red stars. The composite index, which is the scaled sum of the five biological aging indicators, clearly shows most associations and often largest effect sizes. Biological aging was used as outcome in the linear models. Beta for telomere length was multiplied by −1 to compare with other biological clocks. Red stars indicate FDR < 5%. All biological aging indicators were age-regressed, only telomere length was not.

Table 2
Associations between five biological aging indicators and multiple health determinants.

For each biological aging indicator, linear models were fit with the health determinant as predictor, while controlling for sex. Beta’s and p-values from these models are presented here. In the 653 samples with all five data layers available, a composite index was constructed which was significantly associated with more variables than any of the five individual biological aging indicators. All biological aging indicators were age-regressed, only telomere length was not. Telomere length models were corrected for age instead. * Beta for telomere length was multiplied by −1 to compare with other biological aging indicators. All measures are coded such that higher values indicate advanced biological aging. Bold indicates FDR < 5%.

Telomere LengthN=2936EpigeneticclockN=1130Transcriptomic ClockN = 1990Proteomic ClockN=1837Metabolomic ClockN=2910Composite Index (sum)N=653Composite Index (PC1)N=653
Beta*PBetaPBetaPBetaPBetaPBetaPBetaP
DemographicSex (male/female)-0.062.89E-04-0.104.65E-04-0.153.64E-11-0.031.46E-01-0.081.25E-05-0.182.33E-06-0.113.59E-03
Education (# years)-0.031.12E-01-0.025.21E-01-0.016.37E-01-0.053.43E-02-0.038.22E-02-0.043.11E-01-0.052.27E-01
LifestyleAlcohol use (units per week)0.031.05E-01-0.051.40E-010.009.21E-010.072.89E-030.044.57E-020.076.05E-020.091.50E-02
Smoking (pack years)0.063.11E-030.026.22E-010.051.55E-020.101.33E-050.055.09E-030.101.15E-020.122.85E-03
Physical activity0.022.75E-01-0.063.88E-02-0.046.42E-020.031.51E-010.015.18E-01-0.043.62E-010.017.38E-01
Somatic HealthBMI0.041.80E-020.093.94E-030.146.02E-100.129.82E-080.232.07E-350.242.32E-100.222.18E-09
Physical disability0.039.11E-020.111.41E-040.048.61E-020.047.42E-02-0.014.24E-010.107.38E-030.034.01E-01
Lung capacity0.024.19E-010.034.65E-010.042.13E-01-0.041.51E-010.032.37E-010.035.34E-01-0.026.57E-01
Hand grip strength-0.023.33E-01-0.061.71E-010.033.52E-010.017.30E-010.032.24E-01-0.036.14E-010.036.20E-01
Cardiometabolic disease (no/yes)0.023.37E-010.041.56E-010.031.44E-010.031.35E-010.053.94E-030.101.37E-020.083.19E-02
Respiratory disease (no/yes)-0.022.12E-01-0.016.34E-010.022.85E-010.031.27E-010.014.67E-01-0.034.70E-010.017.17E-01
Musculoskeletal disease (no/yes)0.008.11E-01-0.017.37E-010.041.04E-010.024.36E-010.022.23E-010.092.27E-020.114.96E-03
Digestive disease (no/yes)0.035.77E-02-0.025.71E-010.069.76E-030.061.21E-020.022.81E-010.052.01E-010.042.86E-01
Neurological disease (no/yes)-0.022.58E-010.025.60E-010.015.44E-010.022.84E-010.021.93E-01-0.042.64E-01-0.025.09E-01
Endocrine disease (no/yes)-0.014.45E-010.018.13E-01-0.015.75E-010.061.03E-020.031.23E-010.061.18E-010.091.64E-02
Cancer (no/yes)0.009.66E-010.025.65E-010.024.88E-010.031.81E-010.022.01E-010.083.22E-020.075.00E-02
Metabolic syndrome (# components)0.066.35E-040.041.46E-010.139.98E-090.135.34E-090.214.53E-290.289.10E-130.266.41E-12
# Chronic diseases0.007.99E-010.033.63E-010.053.20E-020.091.24E-040.031.39E-010.061.26E-010.078.43E-02
Mental HealthCurrent MDD (no/yes)0.031.59E-010.091.99E-030.071.68E-020.087.62E-03-0.031.61E-010.116.05E-03-0.122.29E-01
Depression severity0.042.40E-020.128.67E-050.032.76E-010.075.99E-03-0.023.74E-010.137.61E-040.051.87E-01
Childhood Trauma0.014.54E-010.127.99E-050.032.06E-010.048.96E-020.042.46E-020.091.96E-020.077.19E-02
  1. * Beta for telomere length was multiplied by -1 to compare with other biological clocks.

    Bold indicates FDR<5%.

Among the lifestyle determinants, alcohol use was associated with advanced proteomic aging (p=3e-3) and smoking (packs per year) was associated with shorter telomere length (p=3e-3), and advanced transcriptomic (p=2e-2), proteomic (p=1e-5) and metabolomic aging (p=5e-3). Physical activity was not associated with any biological aging indicator.

From the somatic health determinants, high BMI was strongly associated with advanced biological aging of all indicators (p=2e-2 for telomere length, p=4e-3 for epigenetic aging, p=6e-10 for transcriptomic aging, p=1e-7 for proteomic aging, and p=2e-35 for metabolomic aging). Physical disability was associated with advanced epigenetic aging (p=1e-4). Within the domain of chronic diseases, the presence of digestive diseases and endocrine diseases were associated with advanced proteomic aging (p=2e-2 and p=1e-2, respectively). Subjects with cardiometabolic disease showed advanced metabolomic aging (p=4e-3) and subjects with digestive disease exhibited advanced transcriptomic aging (p=1e-2). Those with metabolic syndrome showed advanced biological aging across four indicators (p=6e-4 for telomere length, p=1e-8 for transcriptomic aging, p=5e-9 for proteomic aging, p=5e-29 for metabolomic aging).

The presence of current depression and depression severity were associated advanced epigenetic (p=2e-3 and p=9e-5) and proteomic aging (p=8e-3 and p=6e-3, respectively). Current depression was also associated with advanced transcriptomic aging (p=2e-2) and those with childhood trauma showed advanced epigenetic aging (p=8e-5). To verify if the results were confounded by medication use, we computed associations between antidepressant medication (SSRIs, TCAs, or other antidepressants), metabolic-syndrome-related medication (‘metabolic medication’: anti-diabetic, fibrates, or anti-hypertensives) and biological aging (Supplementary file 1). After FDR correction, we found that metabolomic aging was associated with the increased use of metabolic medication (Beta = 0.153, p=2.35e-3), and antidepressant use with proteomic (Beta = 0.208, p=7.16e-5) and transcriptomic aging (Beta = 0.129, p=8.1e-3). The design of the current observational study cannot conclusively prove whether this is a direct medication effect or confounding by indication.

Association between biological aging indicators and mortality in longitudinal analysis

We conducted post-hoc analyses on the relationship between the biological aging indicators and subsequent outcomes after 6 years of follow-up duration. Mortality data and self-reported somatic disease onset (in the categories cardiometabolic, respiratory, musculoskeletal, digestive, neurological, and endocrine diseases, and cancer) was gathered at each measurement wave. There were no significant associations between chronic disease onset or mortality and baseline biological aging, likely due to the low numbers of mortality and disease onset (Supplementary file 3).

Associations between the biological aging composite index and all health determinants

The composite index was computed as the sum of the five scaled biological aging indicators in the 653 samples with data of all five biological levels. Correlations between the five biological aging indicators and the composite index were between 0.43 and 0.51. We found more and stronger associations for the composite index than for any of the individual biological aging indicators: including sex (p=2e-6), BMI (p=2e-10), smoking (p=2e-2), metabolic syndrome (p=9e-13), current MDD (p=6e-3), depression severity (p=7e-4), and childhood trauma (p=2e-2). As an alternative approach, Principal Component Analysis (PCA) was used to compute an alternative composite index. We used the first principle component (PC) of this analysis, which was a weighted sum of the biological aging indicators (for telomere length the weight (w) = 0.042, epigenetic aging w = 0.094, transcriptomic aging w = 0.220, proteomic aging w = 0.707, metabolomic aging w = 0.664), reflecting the highest correlations between the biological aging indicators, which is between metabolomic and proteomic aging. Compared to the composite index that was based on the sum and thus gives equal weight to all five biological aging indicators, the PC-based index had less significant associations with sex, smoking, BMI, and metabolic syndrome. The PC-based index was not significantly associated with physical disability, or mental health outcomes, as opposed to the summed index. The five PC’s each explain more than 15% of variance (the first 2 PC’s more than 25% each), indicating the multidimensionality and non-redundancy of the five biological aging indicators.

To allow for direct effect size comparisons between the composite (summed) index and the individual aging indicators, we compared the findings for the composite index to those of each individual biological aging indicator with the same subsample. In this analysis, p-values and effect sizes were often more pronounced for the composite index (Figure 4, Supplementary file 2). For example, sex, BMI, metabolic syndrome and current MDD, were significantly associated with the composite index, but the betas for the composite index were larger than the betas from any individual indicator. For the other five variables significantly associated with the composite index (smoking, physical disability, cardiometabolic disease, depression severity, and childhood trauma) the betas for the composite index were larger than four out of five betas from the individual biological aging indicators.

Barplots of betas from associations between biological aging and health determinants.

For each of the associations between biological aging and health determinants, the standardized beta and standard deviation derived from linear models were plotted. Only samples that had data for all five biological clocks (N = 653) were used. All biological aging indicators were age-regressed, only telomere length was not.

Discussion

In this study, we examined five biological clocks based on telomere length and four omics levels from a large, clinically well-characterized cohort. We demonstrated significant intercorrelations between three pairs of biological aging indicators, illustrating the complex and multifactorial processes of biological aging. Furthermore, we observed both overlapping and unique associations between the individual biological aging indicators and different lifestyle, somatic and mental health determinants. Separate linear regressions showed that male sex, high BMI, smoking, and metabolic syndrome were consistently associated with more advanced levels of biological aging across at least four of the biological clocks. Strikingly, depression was associated to more advanced levels of epigenetic, transcriptomic and proteomic aging, signifying that both somatic and mental health are associated with advanced biological aging. Finally, by integrating a composite index of all biological aging indicators we were able to obtain larger effect sizes with for example physical disability and childhood trauma exposure, underscoring the broad impact of determinants on cumulative multi-system biological aging.

The range of correlations among the biological aging indicators considered in this study indicates that the correlates of chronological age in different molecular layers were not strongly correlated, suggesting that biological aging may be differently manifested at certain cellular levels. Consistent with prior studies, we showed weak correlations between different biological aging indicators (Li et al., 2020) and we confirm the absent relationship between telomere length and epigenetic aging (Marioni et al., 2018; Belsky et al., 2017Breitling et al., 2016), but also show lack of associations with transcriptomic, proteomic or metabolomic aging. However, we do confirm an earlier finding showing a significant but modest correlation between epigenetic and transcriptomic aging (Peters et al., 2015). The correlation between metabolomic and proteomic aging may partly be explained by the fact that both data were obtained from platforms that were aimed at probing central inflammation lipid processes, rather than the full proteome or metabolome. Nevertheless, we can infer that only some biological aging indicators show overlap, while most of them seem to be tracking distinctive parts of the aging process, even if they are associated with the same somatic or health determinants.

Our study showed that several of the determinants considered exhibited consistent associations with different biological aging indicators. First, male sex was associated with shorter telomere length and advanced epigenetic, transcriptomic, and metabolomic aging, in line with a large body of literature that shows advanced biological aging and earlier mortality in males compared to females (Austad and Fischer, 2016). Second, high BMI was consistently related to all biological aging indicators, showing that the more overweight or obese, the higher the biological age (Gielen et al., 2018), also after controlling for sex. Earlier studies showed similar associations between high BMI and shorter telomere length (Gielen et al., 2018), and older epigenetic (Horvath and Raj, 2018) and transcriptomic aging signatures (Peters et al., 2015). Third, our analyses showed similarly consistent associations between the prevalence of metabolic syndrome and advanced levels of aging. Further, all but epigenetic aging was advanced with respect to cigarette smoking.

Major depressive disorder (MDD) status was consistently related to advanced aging in three (epigenetic, transcriptomic, proteomic) out of the five biological aging indicators. In contrast, a recent study (N > 1000) in young adults (20–39 years) did not show associations between mental health (as measured by the CIDI) and biological aging (indicated by telomere length, homeostatic dysregulation, Klemer and Doubal’s method and Levine’s method) (Hastings et al., 2019), but it seems possible that this sample was too young to fully develop aging-related manifestations of mental health problems, or lacked age variation. It is likely that our data (obtained from participants 18–64 years) may have been more sensitive in picking up associations with mental health due to increased variation in both chronological age (i.e. inclusion of older persons), as well as symptom severity. To further examine whether the results were consistent across participants with and without depressive psychopathology, we repeated all models in post-hoc analyses and added an interaction term between current depression status and health determinants. There was an overall consistent pattern of non-significant interaction terms for most health determinants and biological aging, although only higher BMI was significantly associated to advanced epigenetic aging in the psychopathology group. However, taken together, the results suggest that findings are not different in persons with and without mental disorders. We observed some significant associations between biological aging and medication use. The design of the current observational study cannot conclusively prove whether this is a direct medication effect or confounding by indication: the patient group using antidepressant medication is also the group that is more chronically and severely depressed. This is similar for the metabolic syndrome related medication. Future studies using randomized clinical trial designs are needed to investigate the mechanism of action of direct pharmacological effects of medication on biological aging.

Furthermore, we computed a composite index by summing up the five biological aging indicators studied here. In other words, this integrative metric contains cumulative independent signal from the individual markers and dependent shared signal – with possible reduced noise due to the summation – between them. Given that this composite index demonstrates larger effect sizes for BMI, sex, smoking, depression severity, and metabolic syndrome than the individual aging indicators, it is suggested that being biologically old at multiple cellular levels has a cumulative multi-systemic effect. When integrated, the composite index reveals stronger (i.e. greater cumulative betas for the composite index than individual clocks) converging associations with sex, BMI, metabolic syndrome and current MDD. This provides further support for the hypothesis that not one biological clock sufficiently captures the biological aging process and that not all clocks are under the control of one unitary aging process. There is abundant room for further progress in determining whether biological aging can be modified by intervening on these determinants.

Nonetheless, the question remains which biological mechanism could plausibly link the current quantification of biological aging and its lifestyle, somatic, and mental health determinants. Part of this answer requires discussion on the features used to build the different clocks: the proteomic and metabolomic clocks mostly measure inflammatory or metabolic factors, two highly integrated processes in aging and aging-related diseases (Frasca et al., 2017). Previous studies suggest immune-mediated mechanisms (specifically inflammatory signaling) connecting metabolic syndrome (Révész et al., 2015), mental health disorders (Wohleb et al., 2016), and aging (Révész et al., 2018). Moreover, MDD is a condition in which inflammation, obesity, and premature or advanced aging co-occur and converge. It might therefore be speculated that immunity and 'inflammaging' (Franceschi et al., 2018) may tie together the currently observed associations.

This study did not include existing biological clocks. While the application of established algorithms would increase generalizability of our findings, there are several reasons why it was not optimal to implement previously published algorithms in the NESDA data. First and foremost, generated omics data are platform-dependent and the existing epigenetic (Horvath, 2013) and gene expression (Peters et al., 2015) clocks rely on arrays with different coverage of probes as was used in NESDA, that also target different parts of genes. Second, a subsample of NESDA was part of the previously published metabolomic clock (Akker et al., 2019), thus application of this model to the current dataset would result in overfitting. The current proteomic platform has not been used before to train a biological clock. Moreover, there is currently no validated gold standard for calculating transcriptomic, proteomic, or metabolomic clocks. Importantly, in spite of these limitations, we have followed an alternative but consistent methodological approach for training our omics-based biological clocks, leveraging the advantage to compare, combine, and integrate these clocks within the same population. However, we emphasize the need for epidemiological replication of these determinants in other datasets (e.g. those including different ethnicities) and we recognize that data harmonization and pooling are important strategies on the scientific research agenda that may overcome this limitation in the future.

Since no previously published algorithms were used, we trained our own clocks using ridge regression with cross-validation. This approach relies on the assumption that the determined cross-sectional correlation between the omics patterns and chronological age arise mainly as a consequence of biological aging, and is independent from potential secular trends (Nelson et al., 2020; Belsky, 2015; Belsky et al., 2020). As common to cross-sectional studies, it is, however, impossible to completely rule out potential cohort effects or uncontrolled individual differences and results should be interpreted in light of this limitation. Future longitudinal studies are needed to identify patterns of biological changes that go beyond their ability to predict age at the time of sampling. While the current study only used chronological age as criterion endpoint, it is important to mention that other epigenetic clocks exist that are trained to predict other potential criteria such as phenotypic markers of age (DNAm PhenoAge) (Levine et al., 2018) or a composite biomarker that was derived from DNAm surrogates and smoking in pack-years (GrimAge) (Hillary et al., 2019). Such clocks were developed to lead to improved predictions of risk of mortality.

More research is needed to elucidate whether: (1) physiological disturbances, such as loss of inflammatory control associated with somatic and psychopathology, accelerate biological aging over time, (2) advanced biological aging precedes and constitutes a vulnerability factor that causes somatic and psychopathology, or (3) somatic and psychopathology and biological aging processes are not causally linked, but share underlying etiological roots (e.g. shared genetic risks or environmental factors) (Han et al., 2019). Yet, it could conceivably be hypothesized that dysregulation of immunoinflammatory control may be related to metabolic outcomes, aging, and depression (Diniz and Vieira, 2018), providing scope as to why some of these determinants converge across different platforms and multiple biological levels.

Here, we used a large cohort that was well-characterized in terms of demographics, lifestyle, and both somatic and mental health assessments, to study and integrate five biological clocks across multiple levels of analysis. This is particularly important as we show that the determinants of biological aging encompass several different domains. Moreover, our sample was adequately powered to detect statistically significant associations, limiting the possibility for chance findings and increasing probability for identifying robust biological age determinants. On the other hand, an obvious limitation is the cross-sectional nature of this study that prevents us from drawing any conclusions on whether the determinants accelerate the aging trajectory over time, the other way around, or whether ‘third’ variables effect this association.

Another aspect that limits the interpretability of our findings in the context of increased risk of developing aging-related diseases and mortality was the relatively young age of the current sample. To illustrate, we were unable to predict future incidence of chronic disease or mortality from baseline biological aging, likely due to the low numbers of mortality and disease onset (Supplementary file 3), for example the number of deceased cases ranged from 64 (TL) to 27 (proteomic clock). Previous studies that have associated biological aging with mortality risk commonly include aging cohorts (Danish longitudinal twin study with mean age of 86.1 years; Framingham Offspring Study with mean age 61.0 years; Swedish population cohort SATSA with mean age 63.6 years; German population cohort ESTHER with mean age 62.5 years; Lothian Birth Cohorts with mean age >69.5 years; Normative Aging Study with mean age 71.7 years) (Marioni et al., 2018; Li et al., 2020; Christiansen et al., 2016; Perna et al., 2016; Murabito et al., 2018; Chen et al., 2016). Before definitively interpreting a ‘clock’ as a measure of biological aging, further independent studies are needed to establish that the clock changes with advancing age and forecasts disease, disability and mortality.

Conclusions

In conclusion, this study examined the overlap between five biological aging indicators and their shared and unique associations with somatic and mental health. Our findings indicate that they largely track distinct, but also partially overlapping aspects of this aging process. Further, we demonstrated that male sex, smoking, higher BMI and metabolic syndrome were consistently related to advanced aging at multiple biological levels. Remarkably, our study also converges evidence of depression and childhood trauma associations across multiple platforms, cellular levels, and sample sizes, highlighting the important link between mental health and biological aging. Taken together, our findings contribute to the understanding and identification of biological age determinants, important to the development of end points for clinical and epidemiological research.

Materials and methods

Study design and participants

Request a detailed protocol

Data used were from the Netherlands Study of Depression and Anxiety (NESDA), an ongoing longitudinal cohort study examining course and consequences of depressive and anxiety disorders. The NESDA sample consists of 2981 persons between 18 and 65 years including persons with a current or remitted diagnosis of a depressive and/or anxiety disorder (74%) and healthy controls (26%). Individuals were recruited from mental health care settings, general practitioners, and the general population in the period from September 2004 to February 2007. Persons with insufficient command of the Dutch language or a primary clinical diagnosis of other severe mental disorders, such as severe substance use disorder or a psychotic disorder were excluded. Participants were assessed during a 4 hr clinical visit, consisting of the collection of all somatic and mental health determinants in the current study, as well as a fasting blood draw. All omics data was obtained from the same blood sample, drawn at the same time point as the health determinant examination during the face-to-face visit. The study was approved by the Ethical Review Boards of participating centers, and all participants signed informed consent. More than 94% of the NESDA participants were from North European origin. The population and methods of the NESDA study have been described in more detail elsewhere (NESDA Research Consortium et al., 2008).

Data to derive different biological clocks was available for different subsamples and all based on the same fasting blood draw from participants in the morning between 8:30 and 9:30 after which samples were stored in a −80°C freezer or – for RNA - transferred into PAXgene tubes (Qiagen, Valencia, California, USA) and stored at −20°C. To create biological clocks, we used telomere length (N = 2936), DNA methylation (N = 1130, MBD-seq, 28M CpGs), gene expression (N = 1990, Affymetrix U219 micro arrays, >20K genes), proteins (N = 1837, Myriad RBM DiscoveryMAP 250+, 171 proteins) and metabolites (N = 2910, Nigthingale platform, 231 metabolites), see Table 1 and details in the following sections.

Biological clock assessments

Telomere length

Request a detailed protocol

Leukocyte telomere length was determined at the laboratory of Telomere Diagnostics, Inc (Menlo Park, CA, USA), using quantitative polymerase chain reaction (qPCR), adapted from the published original method by Cawthon, 2002. Telomere sequence copy number in each patient’s sample (T) was compared to a single-copy gene copy number (S), relative to a reference sample. The resulting T/S ratio is proportional to mean leukocyte telomere length. The detailed method is described elsewhere (Verhoeven et al., 2014). The reliability of the assay was adequate: eight included quality control DNA samples on each PCR run illustrated a small intra-assay coefficient of variation (CV = 5.1%), and inter-assay CV was also sufficiently low (CV = 4.6%).

DNA methylation (epigenetic clock)

Request a detailed protocol

To assay the methylation levels of the approximately 28 million common CpG sites in the human genome, we used an optimized protocol for MBD-seq (Han et al., 2018; Aberg et al., 2020). With this method, genomic DNA is first fragmented and the methylated fragments are then bound to the MBD2 protein that has high affinity for methylated DNA. The non-methylated fraction is washed away and only the methylation-enriched fraction is sequenced. This optimized protocol assesses about 94% of the CpGs in the methylome. The sequenced reads were aligned to the reference genome (build hg19/GRCh37) with Bowtie2 (Langmead and Salzberg, 2012) using local and gapped alignment. Aligned reads were further processed using the RaMWAS Bioconductor package (Shabalin et al., 2018) to perform quality control and calculate methylation scores for each CpG.

Gene expression (transcriptomic clock)

Request a detailed protocol

RNA processing and assaying -done at Rutgers University Cell and DNA repository- have been described previously (Jansen et al., 2014; Jansen et al., 2017; Wright et al., 2014). Samples were hybridized to Affymetrix U219 arrays (Affymetrix, Santa Clara, CA). Array hybridization, washing, staining, and scanning were carried out in an Affymetrix GeneTitan System per the manufacturer’s protocol. Gene expression data were required to pass standard Affymetrix QC metrics (Affymetrix expression console) before further analysis. We excluded from further analysis probes that did not map uniquely to the hg19 (Genome Reference Consortium Human Build 37) reference genome sequence, as well as probes targeting a messenger RNA (mRNA) molecule resulting from transcription of a DNA sequence containing a single nucleotide polymorphism (based on the dbSNP137 common database). After this filtering step, data for analysis remained for 423,201 probes, which was summarized into 44,241 probe sets targeting 18,238 genes. Normalized probe set expression values were obtained using Robust Multi-array Average (RMA) normalization as implemented in the Affymetrix Power Tools software (APT, version 1.12.0, Affymetrix). Data for samples that displayed a low average Pearson correlation with the probe set expression values of other samples, and samples with incorrect sex-chromosome expression were removed.

Proteins (proteomic clock)

Request a detailed protocol

As described previously (Lamers et al., 2016), a panel of 243 analytes (Myriad RBM DiscoveryMAP 250+) involved in various hormonal, immunological, and metabolic pathways was assessed in serum using multiplexed immunoassays in a Clinical Laboratory Improvement Amendments (CLIA)-certified laboratory (Myriad RBM; Austin, TX, USA). After excluding analytes with more than 30% missing data (mostly due to values outside the ranges of detection), 171 of the 243 analytes remained for analysis (with values below and above detection limits imputed with the detection limit values).

Metabolites (metabolomic clock)

Request a detailed protocol

Metabolite measurements have been described in detail previously (Akker et al., 2019; BBMRI-NL Metabolomics Consortium et al., 2020). In short, a total of 232 metabolites or metabolite ratios were reliably quantified from Ethylenediaminetetraacetic acid plasma samples using targeted high-throughput proton Nuclear Magnetic Resonance (1H-NMR) metabolomics (Nightingale Health Ltd, Helsinki, Finland) (Soininen et al., 2015). Metabolites measures provided by the platform include (1) lipids, fatty acids and low-molecular-weight metabolites (N = 51); (2) lipid composition and particle concentration measures of lipoprotein subclasses (N = 98); (3) metabolite ratios (N = 81). This metabolomics platform has been extensively used in large-scaled epidemiological studies in the field of diabetes, cardiovascular disease, mortality and alcohol intake (Akker et al., 2019; Würtz et al., 2016; Wurtz et al., 2012; Würtz et al., 2015; Fischer et al., 2014). The data contained missing values due to detection limits. Samples with more than 25 missings were removed (N = 71), metabolites with more than 250 missings were removed (N = 1). Other missing values were replaced with the median value per metabolite. In total 231 metabolites in 2910 samples remained for analysis.

Building biological clocks for multiple omics domains

Request a detailed protocol

Telomere length was multiplied by −1 to be able to compare directions of effects consistent with that of other biological clocks. For each of the other four omics domains (epigenetic, transcriptomic, metabolomic, and proteomic data) the same approach was used to compute biological clocks. First, the omics data were residualized with respect to technical covariates (batch, lab). Second, data per omics marker were normalized using a quantile-normal transformation. Finally, biological age was computed using cross-validation by splitting the sample in 10 equal parts. For each of the 10 groups, nine parts were used as training set and the 10th as test set. In the training set the biological age estimator was computed using ridge regression (R library glmnet), with chronological age as the outcome, and the omics data as predictors. Only for methylation and gene expression a selection of predictors (CpGs for methylation-based models and genes for gene-expression-based models) was made for each cross validation step: we increased the number of sites included in the elastic net in steps (steps for CpGs: 0, 100, 1000, 10,000, 80,000, 100,000, steps for gene expression 100, 500, 1000, 1200, 1400). CpGs/genes were selected in the order of their ranks derived from the association with age in the training sample. We selected the number of CpGs/genes where the cumulative association signal reached a stable plateau. This approach is based on the rationale that adding more markers should theoretically never decrease predictive power. We previously performed tests where the number of CpGs/genes was included in the loop over the k-folds. However, as it produced very similar results but is much more computer intensive (Clark et al., 2020), this latter approach was not used. This approach resulted in 80,000 CpGs (mapping to 2976 genes) for the epigenetic clock, and 1200 probes (mapping to 767 genes) for the transcriptomic clock. For the proteomic and metabolomic data, all markers were used to predict age, because leaving markers out decreased the prediction accuracy. The predictor was then used in the test set to create an unbiased omics-based biological age. For each omics domain, biological aging was defined as the residuals of regressing biological age on chronological age (Han et al., 2018; Peters et al., 2015). Thus, in the terminology we use here, the biological aging indicators represent the biological age acceleration: a positive value means that the biological age is larger than the chronological age. A composite index of biological aging was made by scaling each of the five biological indicators and taking the sum, in the 653 samples that had data for all five omics levels.

Health determinants

Lifestyle

Request a detailed protocol

Alcohol consumption was assessed as units per week by using the AUDIT (Babor et al., 1989). Smoking status was assessed by pack years (smoking duration * cigarettes per day/20). Physical activity (Gerrits et al., 2013) was assessed using the International Physical Activity Questionnaire (IPAQ) (Craig et al., 2003) and expressed as overall energy expenditure in Metabolic Equivalent Total (MET) minutes per week (MET level * minutes of activity * events per week).

Somatic health

Request a detailed protocol

BMI was calculated as measured weight divided by height-squared. Functional status is one of the most potent health status indicators in predicting adverse outcomes in aging populations (Guralnik et al., 1996), including depression (Milaneschi and Penninx, 2014). Assessment of functional status includes measures of physical impairments and disability, reflecting how individuals’ limitations interact with the demands of the environment. Two measures of physical impairments were available: Lung capacity was determined by measuring the peak expiratory flow (PEF in liter/minute) using a mini Wright peak flow meter. Hand grip strength was measured with a Jamar hand held dynamometer in kilograms of force and was assessed for the dominant hand. Furthermore, physical disability was measured with the World Health Organization Disability Assessment Schedule II (WHODAS-II)s the sum of scale 2 (mobility) and scale 3 (self-care). The number of self-reported current somatic diseases for which participants received medical treatment was counted. We used somatic disease categories as categorized previously (Gerrits et al., 2013; Gaspersz et al., 2018): cardiometabolic, respiratory, musculoskeletal, digestive, neurological and endocrine diseases, and cancer. Metabolic syndrome components included waist circumference, systolic blood pressure, HDL cholesterol, triglycerides, and glucose levels, which measurement methods are described elsewhere (Révész et al., 2014).

Mental health

Request a detailed protocol

Presence of current (6 month recency) major depressive disorder was assessed by the DSM-IV Composite International Diagnostic Interview (CIDI) version 2.1. Depressive severity levels in the week prior to assessment were measured with the 28-item Inventory of Depressive Symptomatology (IDS) self-report (Rush et al., 1996). Childhood trauma was assessed with the Childhood Trauma Interview (CTI) (de Graaf et al., 2002). In this interview, participants were asked whether they were emotionally neglected, psychologically abused, physically abused or sexually abused before the age of 16. The CTI reports the sum of the categories that were scored from 0 to 2 (0: never happened; 1: sometimes; 2: happened regularly), which was categorized into five categories.

Statistical analyses

Request a detailed protocol

For each of the five biological aging indicators we computed associations with demographic (sex, education), lifestyle (physical activity, smoking, alcohol use), somatic health (BMI, hand grip strength, lung function, physical disability, chronic diseases), and mental health (current depression, depression severity, childhood trauma) determinants using linear models with health determinants as predictors and biological aging as outcome (for each health determinant separately). All models included a covariate for sex, except for when sex was the outcome. For telomere length, chronological age was used as covariate in the models, for the other biological aging indicators age was not used as covariate because they are independent of chronological age by design. Standardized betas from these models are reported (by scaling predictor and outcome). Correction for multiple testing was done using permutation based FDR (Fehrmann et al., 2011). Subject labels were permuted 1000 times and associations were computed using the permuted data (all biological aging indicators vs all health determinants). For each of the observed p-values (p) the FDR was computed as the average number of permuted p-values smaller than p, divided by the amount of real p-values smaller than p, resulting in a p-value threshold of 2e-2 for a FDR of 5% for all tests. In the 653 overlapping samples with data in each biological clock domain, we scaled (mean 0, standard deviation 1) and summed up the five biological aging indicators in order to create a composite index of biological aging.

Longitudinal analysis of mortality and chronic disease onset

Request a detailed protocol

As NESDA is a longitudinal study, with several follow-up measurement waves, we conducted post-hoc analyses on the relationship between the biological aging indicators and subsequent outcomes after six years of follow-up duration. The average chronological age of our cohort (mean = 41 years, sd = 13, range = 18–65 years) is rather young, so high rates of mortality and morbidity were not expected. Mortality data was gathered at each measurement wave. Also, at each wave self-reported somatic diseases for which participants received medical treatment were assessed. Based on this, we created somatic disease categories as categorized previously (Gerrits et al., 2013; Gaspersz et al., 2018): cardiometabolic, respiratory, musculoskeletal, digestive, neurological and endocrine diseases, and cancer. For these categories, we computed chronic disease onset defined as the disease not being present at baseline (time of biological aging assessment) and present at the latest wave (6 years after baseline). For each biological clock, we computed longitudinal analyses, using a linear model with mortality or chronic disease onset as outcome, and the biological clock residualized for chronological age as predictor, while correcting for sex.

References

Decision letter

  1. Jessica K Tyler
    Senior Editor; Weill Cornell Medicine, United States
  2. Sara Hägg
    Reviewing Editor; Karolinska Institutet, Sweden
  3. Sara Hägg
    Reviewer; Karolinska Institutet, Sweden
  4. Daniel WW Belsky
    Reviewer; Columbia University Mailman School of Public Health, United States
  5. Erik van den Akker
    Reviewer
  6. David G Le Couteur
    Reviewer; University of Sydney, Australia

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

The paper by Jansen et al. investigates somatic and mental health using molecular substrates from the same individuals to create five different multi-omics biological aging clocks. This is exactly the kind of data needed to advance the field and to understand how different layers of data can be integrated to understand biological aging processes.

Decision letter after peer review:

Thank you for submitting your article "An integrative study of five biological clocks in somatic and mental health" for consideration by eLife. Your article has been reviewed by four peer reviewers, including Sara Hägg as the Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Jessica Tyler as the Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Daniel W Belsky (Reviewer #2); Erik van den Akker (Reviewer #3); David G Le Couteur (Reviewer #4).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

In this article, Jansen et al. take some steps toward addressing a knowledge gap on different molecular data used to derive biological aging clocks. They train algorithms to predict chronological age based on several molecular substrates assayed from blood samples: DNA methylation, gene expression, metabolomics, proteomics, and also telomere length. They then test correlations among the several derived measures and compare their associations with various exposures and criterion endpoints relevant to the aging process. The main finding is that the age-correlated features of the different substrates are (a) not very well correlated with one another and (b) largely non-overlapping in their information about health and exposure history. The authors also show that combining information across substrates produces a superior measurement as compared to substrate-specific measures.

Revisions for this paper:

1) Using established clocks

The authors talk about the problem of using their results in other studies since the omics platforms are not always available in other cohorts. It is not clear why it was necessary to train the algorithms in the NESDA cohort. There are published "clocks" for all of the substrates analyzed here (many examples are cited in the Introduction). For example, is there any way that the authors can calculate some of the standard epigenetic clocks (Horvath, Hannum, PhenoAge, GrimAge etc) perhaps not using the online calculator but using standalone scripts for these clocks adapted to your data format? Then, the results would be much more interesting for a wider audience and generalizable to other studies as well. The manuscript would be substantially stronger if these clocks were used in place of bespoke versions derived from the data used for testing hypotheses. If it is not feasible to implement published clocks in the NESDA data, this needs to be explained to the reader.

2) Comments on the algorithms used

If established clocks cannot be used, the alternative strategy of training and testing clocks within a single dataset needs to be presented as the alternative along with specific acknowledgement of the limitations of this approach. The Ridge-regression method used to train the clocks requires the assumption that patterning of molecular markers across the chronological age distribution in the sample reflects biological changes that occur with aging. That assumption has important limitations in any sample, for example see Nelson et al., 2020 on mortality selection or discussion of cohort effects in Belsky et al., 2015 or 2020 . But, depending on how age relates to sampling in the NESDA, there could be further challenges here.

Please also specify whether the feature selection for CpGs/genes was done on the whole dataset, prior to cross-validation, or within the cross-validation loop. If the first, this would lead to reporting overoptimistic performances (overtraining), if the latter, OK; please state so in the manuscript. Please also indicate what step size was used.

Would Mahalanobis distance be a better/more interesting way of analysing the data (eg Bello and Dumancas, Curr Aging Sci, 2017)?

A further consideration to be addressed if the NESDA data are to be used in training the "clocks": prediction of chronological age is only one criterion endpoint used to develop biological aging measures. Recent DNA methylation algorithms including the PhenoAge Clock and the GrimAge Clock were developed from analysis of physiology and mortality data along with chronological age. Some acknowledgement is needed that chronological age is only one of several potential criteria on which to train these measures.

Finally, algorithms trained by applying machine learning analysis to fit high-dimensional molecular data to chronological age variation are hypothesized to measure biological processes of aging. But this is a hypothesis, not a fact. Before we can interpret a "clock" as a measure of biological aging, we must establish that it changes with advancing age, forecasts disease, disability, and mortality, and indicates more advanced/delayed aging in individuals with exposure histories linked to shorter/longer healthy lifespan. The authors should be commended for undertaking some of this testing in NESDA, although using established clocks would be a better alternative. Hence, caution is warranted in interpretation of findings.

3) The NESDA cohort

More detail on NESDA is required to help the reader understand its appropriateness as a setting for comparing measures of aging. How was the sample selected? When were biological measurements collected and what was the extent of attrition from the baseline sample at those time points? In addition, it is not clear when the various exposure and health outcome measurements were collected relative to the biological measurements used to compose the clocks. For each participant, were all the analyses done in a blood sample taken at the same time, or were the different methods applied to bloods taken at different times? How long ago were the samples taken? Are there any storage effects that might influence analyses? A figure illustrating the timeline of data collection would greatly improve clarity of the analysis design.

4) The composite index

The composite index of the 5 clocks was a nice addition to the results. Ideally, clocks, including the composite, are scaled prior to associations with outcomes, as effect sizes are directly compared in the paper. Please indicate whether this has been done; if not, evaluations should be made on the basis of significance only, if so, great! Please state so in the Materials and methods.

Would the results have been similar if the 5 clocks were used combined in multivariate models instead? For example, what happens when all five clocks are put in the same model as predictors with BMI as health outcome? Will they all still be important in the association or are some redundant? Alternatively, it might be useful to compare the current operationalization of the multi-substrate composite to one derived from a factor analysis instead. This is information that is useful to understand the stability of the results but now somehow missed using only one composite index.

5) The Results

The biggest concern with the study is the generalizability given that the samples come from a clinical cohort with individuals suffering from depression and anxiety disorders. About 26% of the study participants were healthy controls and should then be representing a more general population. Moreover, age range is 20-65 years, so probably misses the age groups where biological changes of old age become dominant. Different participants were analysed for each biomarker, and there was a full dataset on only a fraction (approx. 1/3) of the participants that had individual tests done. Any data on ethnicities in NESDA? It is important to perform sensitivity analyses in selected groups of NESDA addressing these concerns in all associations and conclude if the effects are similar or changed in any important way. This also needs to be addressed in the Discussion section.

Another issue is the direction of effects, since these samples and associations are based on cross-sectional data, the authors correctly state that no conclusion can be made on the cause and consequence pathways. However, the biological clocks are used as outcomes in the linear regression models, why? For somatic health and chronic conditions, these are often treated as outcomes in the models using biological age as predictors. In Figure 3, intuitively this is interpreted as the health determinants are the outcomes.

If NESDA is a longitudinal cohort collected about 15 years ago, is there no other follow-up data on somatic and mental health that can be used then using biological age as predictor and health as outcome?

The authors should be somewhat more circumspect in interpreting the clocks they derive. A conclusion of the article is that biological aging proceeds differently across different molecular substrates. This takes the derived measures too literally. Instead, the finding is that the correlates of chronological age in different molecular substrates are not very well correlated with one another.

Given metabolic syndrome and depression are often medicated conditions, are there any data on medications, or any suggestion that medications might influence biomarkers?

6) Data availability

The statement on data availability is not good enough. Why can some gene expression data be released but not other data on these individuals? Data that are anonymized (the identifier key is thrown away) are not considered as sensitive data and it should hence be possible to release more data in this manner.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "An integrative study of five biological clocks in somatic and mental health" for further consideration by eLife. Your revised article has been evaluated by Jessica Tyler (Senior Editor) and Sara Hägg (Reviewing Editor).

The current version of the manuscript represents a highly responsive revision that addressed most comments. There are some remaining issues that need to be addressed before acceptance, as outlined below:

– The mortality association should be mentioned already in the result section as an additional analysis.

– The biological aging indicator is not clearly described in all figure legends, if it is the residulized age version or not.

– State the direction of effect for the medication analysis in the result section, not just p-values.

– The metabolomic platform is "Brainshake" in Materials and methods?

https://doi.org/10.7554/eLife.59479.sa1

Author response

Essential revisions:

1) Using established clocks

The authors talk about the problem of using their results in other studies since the omics platforms are not always available in other cohorts. It is not clear why it was necessary to train the algorithms in the NESDA cohort. There are published "clocks" for all of the substrates analyzed here (many examples are cited in the Introduction). For example, is there any way that the authors can calculate some of the standard epigenetic clocks (Horvath, Hannum, PhenoAge, GrimAge etc) perhaps not using the online calculator but using standalone scripts for these clocks adapted to your data format? Then, the results would be much more interesting for a wider audience and generalizable to other studies as well. The manuscript would be substantially stronger if these clocks were used in place of bespoke versions derived from the data used for testing hypotheses. If it is not feasible to implement published clocks in the NESDA data, this needs to be explained to the reader.

We thank the reviewers for this comment and explain below why we do not use standard epigenetic clocks (or other published clocks using the omics levels presented here), which, if possible, would indeed have facilitated interpretation and replication of our results.

It is important to note that commonly used methods for assaying DNA methylation depend on the Illumina arrays, platforms that generate variables indicating the percentage of methylated CpGs (values range from 0 to 1, indicating no methylation to 100% methylation), whereas the current study used optimized Methyl-CpG binding domain sequencing, that generate quantitative scores (e.g. scores ranging from 0-20) representing the number of fragments covering a CpG that is proportional to the level of methylation occurring in its locus. Optimized MBD-seq has a more comprehensive coverage of the methylome (interrogation of 94% instead of 2-4% of all 28 million common CpG sites in blood). Computing existing array-based epigenetic clocks (e.g. Horvath) using the MBD-seq data would be suboptimal compared to computing MBD-seq based clocks because: (1) MBD-seq covers much more CpGs than the array-based methylation, and restricting to array based CpGs means not leveraging that coverage, and (2) using only the CpGs used for the Horvath clock from the MBD-seq method does not generate equivalent predictive power.

As for the other omics-based clocks that we computed: there are existing gene expression based clocks (Peters et al., 2015) however, gene expression measures are also platform dependent. The mentioned existing clock used Illumina arrays, while we generated gene expression data from Affymetrix arrays, which (in many cases) target different parts of the gene and not completely overlapping gene sets. The proteomic clock is based on a selected number of proteins from a particular platform with 172 proteins. To the best of our knowledge, this platform has not been used before to compute a proteomic clock. The metabolomic clock was created using the Nightingale platform: this platform has been used before to compute biological age, and an algorithm to compute biological age was provided (van den Akker et al., 2019). However, NESDA was part of this study (and thus part of the “training set” to compute the algorithm): using this algorithm in NESDA may lead to overfitting. More importantly, the above mentioned transcriptomic and metabolomic clocks have not been validated by other independent studies. For the above outlined reasons, we believe that the previously established clocks are suboptimal and/or not suitable for the current available data. While we agree that this comes at the cost of generalizability of findings, we also ask for the reviewers understanding that applying existing clocks is not optimal with the available data. Importantly, by using one consistent approach for calculating the biological clocks using different omics data, we increase the comparability of clocks within our own study.

We now clarify this and write the following in the Discussion to emphasize the limitations of this strategy:

“This study did not include existing biological clocks. […] However, we emphasize the need for epidemiological replication of these determinants in other datasets (e.g. those including different ethnicities) and we recognize that data harmonization and pooling are important strategies on the scientific research agenda that may overcome this limitation in the future.”

2) Comments on the algorithms used

If established clocks cannot be used, the alternative strategy of training and testing clocks within a single dataset needs to be presented as the alternative along with specific acknowledgement of the limitations of this approach. The Ridge-regression method used to train the clocks requires the assumption that patterning of molecular markers across the chronological age distribution in the sample reflects biological changes that occur with aging. That assumption has important limitations in any sample, for example see Nelson et al., 2020 on mortality selection or discussion of cohort effects in Belsky et al., 2015 or 2020 . But, depending on how age relates to sampling in the NESDA, there could be further challenges here.

We thank the reviewers for the notion of these references, and now write in the Discussion:

“Since no previously published algorithms were used, we trained our own clocks using ridge regression with cross-validation. […] Future longitudinal studies are needed to identify patterns of biological changes that go beyond their ability to predict age at the time of sampling.”

Please also specify whether the feature selection for CpGs/genes was done on the whole dataset, prior to cross-validation, or within the cross-validation loop. If the first, this would lead to reporting overoptimistic performances (overtraining), if the latter, OK; please state so in the manuscript. Please also indicate what step size was used.

We now clarified this in the Materials and methods section:

“We increased the number of sites included in the elastic net in steps (steps for CpGs: 0, 100, 1000, 10 000, 80 000, 100 000, steps for gene expression 100, 500, 1000, 1200, 1400). […] This approach resulted in 80,000 CpGs (mapping to 2,976 genes) for the epigenetic clock, and 1,200 probes (mapping to 767 genes) for the transcriptomic clock.”

Would Mahalanobis distance be a better/more interesting way of analysing the data (eg Bello and Dumancas, Curr Aging Sci, 2017)?

We thank the reviewer for this suggestion, but in order to make our results comparable to the most commonly previously used methods with similar research designs, we choose not to use the Mahalanobis distance.

A further consideration to be addressed if the NESDA data are to be used in training the "clocks": prediction of chronological age is only one criterion endpoint used to develop biological aging measures. Recent DNA methylation algorithms including the PhenoAge Clock and the GrimAge Clock were developed from analysis of physiology and mortality data along with chronological age. Some acknowledgement is needed that chronological age is only one of several potential criteria on which to train these measures.

We now acknowledge this in the Discussion:

“While the current study only used chronological age as criterion endpoint, it is important to mention that other epigenetic clocks exist that are trained to predict other potential criteria such as phenotypic markers of age (DNAm PhenoAge)43 or a composite biomarker that was derived from DNAm surrogates and smoking in pack-years (GrimAge)44. Such clocks were developed to lead to improved predictions of risk of mortality.”

Finally, algorithms trained by applying machine learning analysis to fit high-dimensional molecular data to chronological age variation are hypothesized to measure biological processes of aging. But this is a hypothesis, not a fact. Before we can interpret a "clock" as a measure of biological aging, we must establish that it changes with advancing age, forecasts disease, disability, and mortality, and indicates more advanced/delayed aging in individuals with exposure histories linked to shorter/longer healthy lifespan. The authors should be commended for undertaking some of this testing in NESDA, although using established clocks would be a better alternative. Hence, caution is warranted in interpretation of findings.

The reviewer is correct on these points, it is also our understanding that we cannot conclusively state that the biological clocks considered in the current cross-sectional study in fact reflect biological processes of ongoing aging. We agree that this limitation did not receive enough attention in the previous version of the manuscript. We have to acknowledge that our sample is rather young (average age=41, range 18-65 years). Despite the fact that we have longitudinal data on our respondents, it is to be expected that the power to look at aging-related outcomes such as mortality and disease onset in our sample is still limited. However, we now performed additional analyses in which we associated biological aging with mortality and chronic disease onset outcomes. These analyses showed no significant associations with biological aging indicators. We added these findings to the Results section and revised the text to reflect that caution is warranted in interpretation of our findings.

We now write in the Materials and methods section:

“Longitudinal analysis of mortality and chronic disease onset

As NESDA is a longitudinal study, with several follow-up measurement waves, we conducted post-hoc analyses on the relationship between the biological aging indicators and subsequent outcomes after six years of follow-up duration. […] For each biological clock we computed longitudinal analyses, using a linear model with mortality or chronic disease onset as outcome, and the biological clock residualized for chronological age as predictor, while correcting for sex.”

And further in the Discussion:

“Another aspect that limits the interpretability of our findings in the context of increased risk of developing aging-related diseases and mortality was the relatively young age of the current sample. […] Before definitively interpreting a "clock" as a measure of biological aging, further independent studies are needed to establish that the clock changes with advancing age and forecasts disease, disability and mortality.”

3) The NESDA cohort

More detail on NESDA is required to help the reader understand its appropriateness as a setting for comparing measures of aging. How was the sample selected? When were biological measurements collected and what was the extent of attrition from the baseline sample at those time points? In addition, it is not clear when the various exposure and health outcome measurements were collected relative to the biological measurements used to compose the clocks. For each participant, were all the analyses done in a blood sample taken at the same time, or were the different methods applied to bloods taken at different times? How long ago were the samples taken? Are there any storage effects that might influence analyses? A figure illustrating the timeline of data collection would greatly improve clarity of the analysis design.

We thank the reviewer for their comment to add more detailed information of NESDA in the manuscript. We now point out in the revised description that all health determinants and blood-based biological clocks were collected at the same timepoint, making a figure illustrating the timeline of data collection seem redundant as there is also no attrition to report. However, we now improved the clarity of the analysis design by adding a section on the timing of data collection, storage, and detailed report of the baseline interview. The Materials and methods section now reads:

“Data used were from the Netherlands Study of Depression and Anxiety (NESDA), an ongoing longitudinal cohort study examining course and consequences of depressive and anxiety disorders. […] Data to derive different biological clocks was available for different subsamples and all based on the same fasting blood draw from participants in the morning between 8:30 and 9:30 after which samples were stored in a -80°C freezer or – for RNA – transferred into PAXgene tubes (Qiagen, Valencia, California, USA) and stored at −20°C. To create biological clocks, we used telomere length (N=2936), DNA methylation (N=1130, MBD-seq, 28M CpGs), gene expression (N=1990, Affymetrix U219 micro arrays, >20K genes), proteins (N=1837, Myriad RBM DiscoveryMAP 250+, 171 proteins) and metabolites (N=2910, Brainshake platform, 231 metabolites), see Table 1 and details in the following sections.”

4) The composite index

The composite index of the 5 clocks was a nice addition to the results. Ideally, clocks, including the composite, are scaled prior to associations with outcomes, as effect sizes are directly compared in the paper. Please indicate whether this has been done; if not, evaluations should be made on the basis of significance only, if so, great! Please state so in the Materials and methods.

We indeed standardized all our clocks prior to association analyses. So, the reported associations are standardized betas that can be compared across individual biological aging indicators. We better indicated this now in the Materials and methods:

“Standardized betas from these models are reported (by scaling predictor and outcome).”

And

“A composite index of biological aging was made by scaling each of the five biological indicators and taking the sum, in the 653 samples that had data for all five omics levels.”

Would the results have been similar if the 5 clocks were used combined in multivariate models instead? For example, what happens when all five clocks are put in the same model as predictors with BMI as health outcome? Will they all still be important in the association or are some redundant? Alternatively, it might be useful to compare the current operationalization of the multi-substrate composite to one derived from a factor analysis instead. This is information that is useful to understand the stability of the results but now somehow missed using only one composite index.

We agree that there are multiple strategies to compute a composite index. As an alternative we now also report the associations between the first PC of a PCA analysis of the 5 biological aging indicators, and the health determinants. This PC reflects the strongest biological aging correlations, which is between metabolomic and proteomic aging. Thus, this PC is a weighted sum of the indicators, all positive weights, but with highest weights for metabolomic and proteomic aging indicators. Therefore, the associations between this PC and the health outcomes, is very similar to the associations between the health determinants and metabolomic and proteomic aging. Moreover, PCA analysis gives the explained variance of multiple independent dimensions derived from the underlying data, and therefore also helps to answer the reviewers question about redundancy of the biological clocks. We added to the Results:

“As an alternative approach, Principal Component Analysis (PCA) was used to compute an alternative composite index. […] The five PC’s each explain more than 15% of variance (the first 2 PC’s more than 25% each), indicating the multidimensionality and non-redundancy of the five biological clocks.”

5) The Results

The biggest concern with the study is the generalizability given that the samples come from a clinical cohort with individuals suffering from depression and anxiety disorders. About 26% of the study participants were healthy controls and should then be representing a more general population. Moreover, age range is 20-65 years, so probably misses the age groups where biological changes of old age become dominant. Different participants were analysed for each biomarker, and there was a full dataset on only a fraction (approx. 1/3) of the participants that had individual tests done. Any data on ethnicities in NESDA? It is important to perform sensitivity analyses in selected groups of NESDA addressing these concerns in all associations and conclude if the effects are similar or changed in any important way. This also needs to be addressed in the Discussion section.

We would like to emphasize that sample size is important for testing robust associations, so we wanted to make optimal use of the current sample sizes to examine individual health determinant associations. The rationale to additionally test all associations in a fully overlapping sample with all omics data (N=653) was to compare whether shared and unique associations would be (much) different, but this analysis showed very similar results to those with a maximum number of samples per omics level, indicating high sensitivity of the results.

Furthermore, although we believe that poorer mental or somatic health outcomes will be similarly associated to advanced aging in both control and patient groups, we can see how associations might be confounded by diagnostic status. We therefore repeated the same separate univariate analyses, but included an additional interaction term between current depression status and health determinant in the models. Overall, these interaction terms were not significant. These findings can be found in the supplement.

We now write in the Discussion:

“To further examine whether the results were consistent across participants with and without depressive psychopathology, we repeated all models in post-hoc analyses and added an interaction term between current depression status and health determinants. There was an overall consistent pattern of non-significant interaction terms for most health determinants and biological aging, although only higher BMI was significantly associated to advanced epigenetic aging in the psychopathology group. However, taken together, the results suggest that findings are not different in persons with and without mental disorders.”

We have also added the following sentence to the Materials and methods:

“More than 94% of the NESDA participants were from North European origin.”

And the following sentence in the Discussion:

“We also emphasize the need for epidemiological replication of these determinants in other datasets (e.g. those including different ethnicities) and we recognize that data harmonization and pooling are important strategies on the scientific research agenda that may overcome this limitation in the future.”

Another issue is the direction of effects, since these samples and associations are based on cross-sectional data, the authors correctly state that no conclusion can be made on the cause and consequence pathways. However, the biological clocks are used as outcomes in the linear regression models, why? For somatic health and chronic conditions, these are often treated as outcomes in the models using biological age as predictors. In Figure 3, intuitively this is interpreted as the health determinants are the outcomes.

In cross sectional analysis, interchanging predictor and outcome will not significantly change the interpretation of associations between these variables in a linear model. Since we wanted to find determinants for biological aging, while correcting for sex, we used biological aging as outcomes. We now clarify in the legend of Figure 3 that biological clocks residualized for age were used as outcomes.

If NESDA is a longitudinal cohort collected about 15 years ago, is there no other follow-up data on somatic and mental health that can be used then using biological age as predictor and health as outcome?

We appreciate this comment and have conducted additional analyses to examine whether biological aging at baseline predicts the presence of somatic disease and mortality rates 6 years later in those who were initially disease-free at baseline. Please see response 2.

The authors should be somewhat more circumspect in interpreting the clocks they derive. A conclusion of the article is that biological aging proceeds differently across different molecular substrates. This takes the derived measures too literally. Instead, the finding is that the correlates of chronological age in different molecular substrates are not very well correlated with one another.

We assume the reviewers refer to this sentence in the Discussion: “Biological aging seems to be differently manifested at certain cellular levels, as suggested by the range of correlations among the biological clocks considered in this study.”

To make clearer that findings from the biological clocks are only suggesting findings for biological aging we changed this sentence into:

“The range of correlations among the biological aging indicators considered in this study indicates that the correlates of chronological age in different molecular layers were not strongly correlated, suggesting that biological aging may be differently manifested at certain cellular levels.”

Given metabolic syndrome and depression are often medicated conditions, are there any data on medications, or any suggestion that medications might influence biomarkers?

We thank the reviewer for this suggestion, and provide additional analyses in which we verified if antidepressant medication (SSRIs, TCAs, or other antidepressants) or metabolic syndrome related medication (anti-diabetic, fibrates, or anti-hypertensives), were associated with the biological clocks.

The findings can be read in the Results section:

“To verify if the results were confounded by medication use, we computed associations between antidepressant medication (SSRIs, TCAs, or other antidepressants), metabolic syndrome related medication (“metabolic medication”: anti-diabetic, fibrates, or anti-hypertensives) and biological aging (Supplementary file 1). After FDR correction, we found that metabolomic aging was associated with the use of metabolic medication (P=2.35e-3), and antidepressant use with proteomic (P=7.16e-5) and transcriptomic aging (P=8.1e-3). The design of the current observational study cannot conclusively prove whether this is a direct medication effect or confounding by indication.”

We believe, however, that these findings are not indicating direct medication effects per se. In fact, these could be interpreted as confirmation of our disease findings. E.g. the patient group using antidepressant medication is also the group that is more chronically and severely depressed. Therefore, using the current data from an observational cohort, we are unable to discern the effects of depression, depression severity, or direct effects of antidepressants. The same is true for the findings of metabolic medication.

We have added the following sentence to the Discussion:

“We observed some significant associations between biological aging and medication use. The design of the current observational study cannot conclusively prove whether this is a direct medication effect or confounding by indication: the patient group using antidepressant medication is also the group that is more chronically and severely depressed. This is similar for the metabolic syndrome related medication. Future studies using randomized clinical trial designs are needed to investigate the mechanism of action of direct pharmacological effects of medication on biological aging.”

6) Data availability

The statement on data availability is not good enough. Why can some gene expression data be released but not other data on these individuals? Data that are anonymized (the identifier key is thrown away) are not considered as sensitive data and it should hence be possible to release more data in this manner.

We have updated our data availability section.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

The current version of the manuscript represents a highly responsive revision that addressed most comments. There are some remaining issues that need to be addressed before acceptance, as outlined below:

– The mortality association should be mentioned already in the result section as an additional analysis.

We now write in the Results section:

“Association between biological aging indicators and mortality in longitudinal analysis

We conducted post-hoc analyses on the relationship between the biological aging indicators and subsequent outcomes after six years of follow-up duration. Mortality data and self-reported somatic disease onset (in the categories cardiometabolic, respiratory, musculoskeletal, digestive, neurological and endocrine diseases, and cancer) was gathered at each measurement wave. There were no significant associations between chronic disease onset or mortality and baseline biological aging, likely due to the low numbers of mortality and disease onset (Supplementary file 3).”

– The biological aging indicator is not clearly described in all figure legends, if it is the residulized age version or not.

We added to the legends of Figure 3 and Figure 4:

“All biological aging indicators were age-regressed, only telomere length was not.”

– State the direction of effect for the medication analysis in the result section, not just p-values.

We now write in the Results section:

“After FDR correction, we found that metabolomic aging was associated with the increased use of metabolic medication (Β=0.153, P=2.35e-3), and antidepressant use with proteomic (Β=0.208, P=7.16e-5) and transcriptomic aging (Β=0.129, P=8.1e-3). The design of the current observational study cannot conclusively prove whether this is a direct medication effect or confounding by indication.”

– The metabolomic platform is "Brainshake" in Materials and methods?

We corrected this to “Nightingale platform” (formerly known as “Brainshake”).

https://doi.org/10.7554/eLife.59479.sa2

Article and author information

Author details

  1. Rick Jansen

    Department of Psychiatry, Amsterdam UMC, Vrije Universiteit Amsterdam, Amsterdam Public Health Research Institute and Amsterdam Neuroscience, Amsterdam, Netherlands
    Contribution
    Formal analysis, Visualization, Methodology, Writing - original draft, Writing - review and editing
    Contributed equally with
    Laura KM Han
    For correspondence
    ri.jansen@ggzingeest.nl
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-3333-6737
  2. Laura KM Han

    Department of Psychiatry, Amsterdam UMC, Vrije Universiteit Amsterdam, Amsterdam Public Health Research Institute and Amsterdam Neuroscience, Amsterdam, Netherlands
    Contribution
    Writing - original draft, Writing - review and editing
    Contributed equally with
    Rick Jansen
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9647-3723
  3. Josine E Verhoeven

    Department of Psychiatry, Amsterdam UMC, Vrije Universiteit Amsterdam, Amsterdam Public Health Research Institute and Amsterdam Neuroscience, Amsterdam, Netherlands
    Contribution
    Writing - original draft, Writing - review and editing
    Competing interests
    No competing interests declared
  4. Karolina A Aberg

    Center for Biomarker Research and Precision Medicine, Virginia Commonwealth University, Richmond, United States
    Contribution
    Writing - review and editing
    Competing interests
    No competing interests declared
  5. Edwin CGJ van den Oord

    Center for Biomarker Research and Precision Medicine, Virginia Commonwealth University, Richmond, United States
    Contribution
    Writing - review and editing
    Competing interests
    No competing interests declared
  6. Yuri Milaneschi

    Department of Psychiatry, Amsterdam UMC, Vrije Universiteit Amsterdam, Amsterdam Public Health Research Institute and Amsterdam Neuroscience, Amsterdam, Netherlands
    Contribution
    Writing - original draft, Writing - review and editing
    Competing interests
    No competing interests declared
  7. Brenda WJH Penninx

    Department of Psychiatry, Amsterdam UMC, Vrije Universiteit Amsterdam, Amsterdam Public Health Research Institute and Amsterdam Neuroscience, Amsterdam, Netherlands
    Contribution
    Writing - original draft, Writing - review and editing
    Competing interests
    has received research funding (not related to the current paper) from Boehringer Ingelheim and Jansen Research.

Funding

No external funding was received for this work.

Acknowledgements

The infrastructure for the NESDA study (http://www.nesda.nl) is funded through the Geestkracht program of the Netherlands Organisation for Health Research and Development (ZonMw, grant number 10-000-1002) and financial contributions by participating universities and mental health care organizations (VU University Medical Center, GGZ inGeest, Leiden University Medical Center, Leiden University, GGZ Rivierduinen, University Medical Center Groningen, University of Groningen, Lentis, GGZ Friesland, GGZ Drenthe, Rob Giel Onderzoekscentrum). Telomere length assaying was supported through a NWO-VICI grant (number 91811602). Methylation sequencing was supported by NIMH grant R01MH099110. Metabolomics data were generated within the framework of the BBMRI Metabolomics Consortium funded by BBMRI-NL, a research infrastructure financed by the Dutch government (NWO, grant nr 184.021.007 and 184033111). Gene expression data were funded by the US National Institute of Mental Health (RC2MH089951).

Ethics

Human subjects: The NESDA study was approved by the Ethical Review Boards of participating centers, and all participants signed informed consent. The population and methods of the NESDA study have been described in more detail elsewhere (Hillary et al., 2019).

Senior Editor

  1. Jessica K Tyler, Weill Cornell Medicine, United States

Reviewing Editor

  1. Sara Hägg, Karolinska Institutet, Sweden

Reviewers

  1. Sara Hägg, Karolinska Institutet, Sweden
  2. Daniel WW Belsky, Columbia University Mailman School of Public Health, United States
  3. Erik van den Akker
  4. David G Le Couteur, University of Sydney, Australia

Publication history

  1. Received: May 29, 2020
  2. Accepted: December 14, 2020
  3. Version of Record published: February 9, 2021 (version 1)

Copyright

© 2021, Jansen et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 1,340
    Page views
  • 157
    Downloads
  • 1
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

  1. Further reading

Further reading

    1. Cell Biology
    2. Developmental Biology
    Neta Erez et al.
    Research Article Updated

    A hallmark of aging is loss of differentiated cell identity. Aged Drosophila midgut differentiated enterocytes (ECs) lose their identity, impairing tissue homeostasis. To discover identity regulators, we performed an RNAi screen targeting ubiquitin-related genes in ECs. Seventeen genes were identified, including the deubiquitinase Non-stop (CG4166). Lineage tracing established that acute loss of Non-stop in young ECs phenocopies aged ECs at cellular and tissue levels. Proteomic analysis unveiled that Non-stop maintains identity as part of a Non-stop identity complex (NIC) containing E(y)2, Sgf11, Cp190, (Mod) mdg4, and Nup98. Non-stop ensured chromatin accessibility, maintaining the EC-gene signature, and protected NIC subunit stability. Upon aging, the levels of Non-stop and NIC subunits declined, distorting the unique organization of the EC nucleus. Maintaining youthful levels of Non-stop in wildtype aged ECs safeguards NIC subunits, nuclear organization, and suppressed aging phenotypes. Thus, Non-stop and NIC, supervise EC identity and protects from premature aging.

    1. Cell Biology
    2. Neuroscience
    Friederike Elisabeth Kohrs et al.
    Tools and Resources

    Rab GTPases are molecular switches that regulate membrane trafficking in all cells. Neurons have particular demands on membrane trafficking and express numerous Rab GTPases of unknown function. Here we report the generation and characterization of molecularly defined null mutants for all 26 rab genes in Drosophila. In flies, all rab genes are expressed in the nervous system where at least half exhibit particularly high levels compared to other tissues. Surprisingly, loss of any of these 13 nervous system-enriched Rabs yielded viable and fertile flies without obvious morphological defects. However, all 13 mutants differentially affected development when challenged with different temperatures, or neuronal function when challenged with continuous stimulation. We identified a synaptic maintenance defect following continuous stimulation for six mutants, including an autophagy-independent role of rab26. The complete mutant collection generated in this study provides a basis for further comprehensive studies of Rab GTPases during development and function in vivo.