DNA Methylation Clocks and Their Predictive Capacity for Aging Phenotypes and Healthspan

Introduction

In contrast to the steady pace of chronological age, the pace of biological age varies among individuals and may predict distinct aspects of aging at different life stages. For example, biological aging later in life reflects the link with disease, morbidity, and mortality, whereas in youth/midlife, it might predict healthspan or cognitive/physical decline (eg, individuals aging more rapidly are less cognitively/physically able).1,2 As chronological age does not sufficiently represent fundamental aging processes, methods to measure biological aging have been developed, which is important for assessing strategies to slow down biological aging and extend healthspan. A reliable measure of biological age could be a factor in predicting disease onset in presymptomatic carriers of causal mutations, and assist in the development of preventive rather than therapeutic strategies.1

Technical breakthroughs have led to the discovery of several molecular markers of aging, including epigenetic biomarkers.3-6 For instance, genome-wide RNA-interference-based screening in Caenorhabditis elegans revealed a conserved epigenetic mechanism, which implicated 59 genes as modulators of age-related behavioural deterioration rate.7 Two neuronal genes were among the most prominent hits: epigenetic reader (BAZ-2) and histone 3 lysine 9 methyltransferase (SET-6), which could accelerate behavioural deterioration by repressing the expression of nuclear-encoded mitochondrial proteins. Importantly, the expression of human orthologues (BAZ2B and EHMT1) in the frontal cortex increases with age and correlates with Alzheimer disease (AD) progression.

Among biomarkers of aging, such as telomere length (TL), metabolomic, transcriptomic and proteomic variations, the most promising are based on the DNA methylation (DNAm) of cytosines at CpG dinucleotides,8 representing one of the key epigenetic mechanisms altering gene expression or splicing. The cumulative assessment of DNAm levels at age-related CpGs could be used as a DNAm clock,6,9-13 which may mirror biological aging (Figure 1). Although some clinical biomarkers outperform DNAm clocks in reflecting morbidity and mortality,14 the advantage of DNAm clocks is their ability to measure either multitissue or cell-/tissue-specific aging. DNAm clocks could help explain why some individuals stay healthy, whereas others develop age-related neurodegenerative diseases.

Several studies support the link between DNAm clocks and biological age. DNAm-age acceleration (difference between DNAm-age and chronological age) was associated with major neurodegenerative diseases, such as AD,15 Parkinson disease (PD),16 Huntington disease17 and amyotrophic lateral sclerosis (ALS).18,19 Similarly, HIV-infected individuals exhibit premature aging based on methylome-wide changes and Horvath DNAm clock.20,21 Furthermore, individuals with Werner or Down syndromes also display accelerated DNAm clocks.22,23 In contrast, DNAm-age in centenarians is on average 9 years younger than their chronological age.24

However, it is mostly unclear what the underlying molecular mechanisms of DNAm clocks are. Do they reflect similar aspects of the aging process? What is their capacity to predict risk of decline before disease onset and therapeutic effectiveness aiming to extend healthspan? Here, we provide an overview of DNAm clocks, including their application in neurodegenerative diseases, potential confounders and gaps in the current knowledge.

Comparison of Different DNAm Clocks

Most DNAm clocks are based on a limited number of CpGs, which are widely distributed throughout the genome and mainly selected from the 27K or 450K BeadChips that have currently been replaced by the more informative EPIC chip covering 850K CpGs (Table 1). Age-related CpGs are overrepresented in the proximity of Polycomb-binding regions and promoters, which are key regulators of gene expression.6,11,13 The little overlap in CpGs between different DNAm clocks25-27 might be the result of several factors (eg, cell-/tissue-specific differences) and explain why some clocks are capable of catching certain outcomes better than others. In addition, the applied arrays might not include the most informative age-related CpGs.6,14 Many DNAm clocks are obtained from linear regression algorithms trained against chronological age6; however, clocks too perfectly trained on chronological age would not contain information on interindividual variability in biological age.5,6 Finally, some DNAm clocks focus on the comparison between DNAm levels and physiological measures of biological age (eg, cholesterol levels), which may be the consequence of confounding factors (eg, obesity) rather than representing aging itself.28

Comparison of Genes in Different DNAm Clocks

Knowledge about aging mechanisms would enable the selection of the most suitable DNAm clocks for a specific task (eg, in a disease-context) and its biological interpretation for future clinical practices.51 For instance, it would be important to establish the age-related expression pattern of genes corresponding to clock-building CpGs (Supplementary Table 1). The CpGs of the 14 DNAm clocks are mapped to 1633 different genes, however the gene-set underlying DNAm GrimAge is unavailable. Based on the NHGRI-EBI Catalog of genome-wide association studies,52 106 of these loci are genetically associated with one of the major neurodegenerative diseases (ALS, AD, PD, or frontotemporal dementia), including genes responsible for Mendelian forms of PD (PARK2 and ATP13A2) or AD (PSEN1), each of which was included in 2 clocks.

Overall, only 173 genes overlapped in at least 2 DNAm clocks; however, some of them were incorporated in up to 6 clocks (KLF14), pointing to their significance in aging mechanisms. Chronological DNAm clocks appear to have more gene-overlap than biological clocks (Figure 2). This is likely because CpGs used in the biological clocks were selected to correlate with multisystem phenotypic measures, whereas CpGs in chronological DNAm clocks were selected to optimize the prediction of chronological age. As expected, the Zhang clock showed the most gene-overlap with other chronological clocks, because it is a near-perfect predictor of chronological age.33 In the assessment of the 9 chronological vs 5 biological clocks, the strongest gene-overlap was observed between Levine (DNAm PhenoAge) and Weidner clocks, which have been associated with several similar traits, including lifestyle factors, TL and some clinical measures.39 In addition, there is some gene-overlap between Levine clock and both Hannum and Horvath clocks, likely accounting for the fact that each of them has shown associations with similar phenotypes (eg, all-cause mortality and chronological age).

Among the 28 top-overlapping genes (incorporated in at least 3 DNAm clocks), 21 genes display a consistent direction of age-related DNAm (Supplementary Table 2). For example, EDARADD is hypomethylated in 5 different DNAm clocks, suggesting it is upregulated with age, while the hypermethylation of KLF14 suggests it is downregulated with age. The greatest discrepancy was presented by the DNAmTL clock by Lu, which displayed the opposite DNAm direction in 4 of 8 genes shared with other clocks. The DNAmTL clock is rather unique, because it is a DNAm-based estimator of TL, and telomere attrition with its associated epigenetic changes is a distinct cellular process of aging.38,54 Hence, the list of 21 genes prioritized for functional studies and eventually for age-related therapies could be extended (Supplementary Table 2). In general, hypermethylation of age-associated genes results in a reduction of their expression; however, the link between DNAm and gene expression is not always straightforward, because elevated DNAm over the gene body can be accompanied by high gene expression.55

Functional annotation and gene-disease associations (using DAVID 6.8)56 revealed that 12 of the 28 top-overlapping genes are linked to neurodegenerative diseases (Supplementary Table 2). The most striking example is KLF14, integrated in 6 DNAm clocks. Age-related hypermethylation of KLF14 has been associated with abnormalities in DNA repair and cell cycle control in familial early-onset AD, and may be an epigenetic mechanism for the hypermethylation of TRIM95 (incorporated in 2 DNAm clocks) contributing to pro-apoptotic signalling in AD.55 Another example is the association between GRIA2 levels and AD brain pathology.57 Moreover, ASPA was associated with age of onset in Canavan disease,58,59 and the expression levels of PDE4 isoforms were linked to AD.60 Also, an association with VGF levels was reported in AD61 and ALS.62,63 Finally, reduced SCAP levels were linked to cognitive decline in an AD model.64 The observation that half of the top-overlapping genes are associated with neurodegenerative disorders reinforces the possibility that DNAm clocks could be used as disease biomarkers in clinical trials of neurodegenerative diseases.

Studies of DNAm Clocks in Neurodegenerative Diseases

Neurodegenerative diseases share common aberrant epigenomic patterns, which target genes involved in many cellular pathways.65 Multiple studies have used specific DNAm changes as disease biomarkers. For instance, increased DNAm in C9orf72 (in response to repeat expansion) leads to a downregulation of C9orf72 expression and correlates with disease onset and duration.66-68 Studies of neurodegenerative diseases revealed a link between clinical outcomes and DNAm-age acceleration, mainly using the Horvath multitissue clock.11

In ALS, studies of discordant identical twins reported that the ALS-affected twin had greater DNAm-age acceleration than the asymptomatic twin.69-71 Furthermore, in C9orf72 ALS patients (n = 46), every 5-year increase in blood-based DNAm-age acceleration was associated with a 3.2-year earlier age at onset.18 A study of genetically unexplained ALS patients (n = 249) also revealed a strong reverse correlation of DNAm-age acceleration with disease onset and survival.19 A 5-year increase in DNAm-age acceleration was linked to a 6.4-year earlier onset using blood collected at ALS diagnosis and 8.5-year earlier onset using CNS tissues, indicating further acceleration of DNAm-age by the end-stage of disease.

In AD, DNAm-age acceleration was associated with Aβ amyloid load and cognitive decline.15 Notably, 32 genes corresponding to clock-building CpGs are involved in the amyloid biological network (Supplementary Table 1). For instance, BACE1 and PSEN1, which encode Aβ generating β- and γ-secretase, respectively, are incorporated in DNAm PhenoAge.14 A study of the age acceleration measures of both Horvath and Hannum clocks revealed an association with several AD risk factors (eg, body mass index and cigarette smoking status).72 In addition, the Horvath measure correlated with cholesterol levels, while Hannum measure correlated with educational level. A recent investigation of the Horvath clock in a unique family with monozygotic triplets, 2 of whom developed AD in their mid-70s, revealed that the triplets’ DNAm-age was 6 to 10 years younger than their chronological age. In contrast, the triplet’s offspring with early-onset AD (at age 50) had a DNAm-age 9 years older than chronological age, indicating accelerated aging.73

Acceleration of the Horvath clock was also linked to PD, and an association was observed between PD and DNAm-based measures of blood cell types (patients showed more granulocytes but fewer helper T and B cells than controls).16 Furthermore, investigation of a PD family with the p.A53E mutation in α-synuclein showed that an earlier onset was accompanied by increased DNAm-age acceleration.74 Of note, the association between DNAm-age measures and disease traits may depend on the selected DNAm clock. In a study of multiple sclerosis, accelerated phenotypic aging was observed using DNAm PhenoAge, whereas the measures of the Horvath or Hannum clock did not differ from chronological age,51 which likely reflects the different aging mechanisms behind each DNAm clock. It illustrates the importance of choosing the proper age-estimator in disease-specific contexts for accurate interpretation of the results.

Overall, DNAm clocks could provide a better understanding of age-related mechanisms underlying neurodegeneration and help with designing an ‘extreme phenotype’ study/trial (eg, early vs late onset; and fast vs slow progression). Furthermore, DNAm modifiers may serve as potential therapeutic targets, which is supported by the observation that several loci included in DNAm clocks are genetically and/or functionally associated with neurodegenerative diseases (Supplementary Tables 1 and 2). Finally, it may be valuable to use several DNAm clocks together to combine their strengths for a specific study, which would require a better understanding of the aging mechanisms behind each clock and confounding factors influencing clock accuracy.

Confounders of DNAm Clock Accuracy

Discussion and Future Directions

A single best measure of biological aging does not exist, although some DNAm clocks outperform others for specific tasks. Each clock is built according to their unique training method, in which the age range, selected tissue(s), platform/statistical methodology, number of CpGs, and samples are of importance. Increasing the number of CpGs leads to a higher accuracy in the prediction of chronological age11,13,32,33; however, it could also cause the loss of informative disease-related CpGs.45 The little overlap in CpGs between the different clocks14 is likely reflective of diverse aging processes. For instance, CpGs in biological DNAm clocks were selected based on their correlation with multisystem phenotypic measures, in contrast to CpGs linked with chronological age. Notably, among the genes corresponding to clock-building CpGs, there are 28 genes incorporated in more than 3 DNAm clocks, most of which show a consistent direction of age-related DNAm in the different clocks (Supplemental Table 2). These genes could be prioritized for future age-related therapies and functional studies, including an evaluation of whether their expression levels influence disease presentation.

The various DNAm clocks reflect different aspects of a multidimensional aging process, including telomere biology, epigenomic instability, transcription, cellular differentiation and senescence.51 Some DNAm changes could be contributed to disease, while others may be a response to age-related decline, suggesting that DNAm clocks reflect a system maintaining epigenetic stability.11 Increasing DNAm variability over time may be the result of an accumulation of DNAm changes through cell division and DNA replication (eg, from errors by DNA-methyltransferases maintaining DNAm patterns on division).93 However, this explanation cannot account for the cell-/tissue-independent DNAm clocks that accurately predict chronological age even in postmitotic tissues (eg, brain). Chronological DNAm clocks could be driven by an increase in entropy, in which the epigenetic landscape smoothens with age and the ticking rate of the clock may reflect the general progression of low-/high-methylated CpGs towards a more intermediate DNAm level.12,13 Furthermore, they may reflect an evolutionarily selected programmed process, as such clocks remarkably accurately predict chronological age and display a link with tissue development; however, these observations are also consistent with the above-mentioned entropic model.6 In contrast, biological DNAm clocks are more likely to measure the variability between individuals in age-related functional decline and disease. Indeed, these clocks seem to be enriched for genes involved in age-related diseases, and the DNAm changes may affect disease progression through the regulation of these genes (eg, the Polycomb target genes in epiTOC).6,34

Importantly, measures of age acceleration could predict the therapeutic efficacy of drugs aiming to slow aging. For instance, a recent pilot study encouraged the use of the GrimAge predictor in a clinical trial intended to regenerate the thymus to prevent signs of immunosenescence.94 The application of DNAm clocks is especially important in neurodegenerative diseases, because aging is the strongest risk factor for all of them. However, it is unknown whether acceleration of DNAm-age is the cause or a consequence of neurodegeneration, which could be addressed by longitudinal studies (eg, assessing DNAm-age before and after disease onset). A longitudinal approach could also help identify more accurate measures of biological aging. For example, a study assessing 18 biomarkers of mortality at age 26, 32, and 38 revealed that worse healthspan outcomes were linked to a faster pace of aging.2 In the future, it would be important to have a longer follow-up period, use different ethnic cohorts and include the most recent biomarkers of age-related decline.1,2,95 However, a longitudinal study is only possible for easily accessible tissues (eg, blood).

In conclusion, each reported DNAm clock has its own strengths and limitations. Getting a better understanding of the aging mechanisms reflected by each clock is crucial for selecting the most suitable DNAm clock in a disease-specific context.45 It could also assist in determining which DNAm clocks may be beneficial to evaluate together in the same experimental setting to capture the complexity of the aging process and improve predictions of health outcomes with age. However, several confounders may affect the accuracy of DNAm clocks, which should be taken into consideration to correctly interpret the link between age predictions or health outcomes. Novel technologies could detect more CpGs in the human methylome to better track biological aging, and longitudinal measurements would be informative in evaluating the rate of epigenetic aging.1,2,96