Epigenetic age prediction

Abstract Advanced age is the main common risk factor for cancer, cardiovascular disease and neurodegeneration. Yet, more is known about the molecular basis of any of these groups of diseases than the changes that accompany ageing itself. Progress in molecular ageing research was slow because the tools predicting whether someone aged slowly or fast (biological age) were unreliable. To understand ageing as a risk factor for disease and to develop interventions, the molecular ageing field needed a quantitative measure; a clock for biological age. Over the past decade, a number of age predictors utilising DNA methylation have been developed, referred to as epigenetic clocks. While they appear to estimate biological age, it remains unclear whether the methylation changes used to train the clocks are a reflection of other underlying cellular or molecular processes, or whether methylation itself is involved in the ageing process. The precise aspects of ageing that the epigenetic clocks capture remain hidden and seem to vary between predictors. Nonetheless, the use of epigenetic clocks has opened the door towards studying biological ageing quantitatively, and new clocks and applications, such as forensics, appear frequently. In this review, we will discuss the range of epigenetic clocks available, their strengths and weaknesses, and their applicability to various scientific queries.

interventions, a biomarker that infers biological age is required, that is a biological indicator that predicts the health and lifespan of an individual better than chronological age (chAge) (Baker & Sprott, 1988).
In the past decade, a large number of age predictors utilising DNAm have been developed. These DNAm age predictors (more commonly known as epigenetic clocks) are created using CpGs that have tractable changes with age. The majority of these clocks are built using penalised regression models (such as elastic net (Zou & Hastie, 2005) or LASSO (Tibshirani, 1997)), which select a group of CpGs that have a monotonically increasing relationship with age in a given training data . In other words, key CpGs whose agerelated hyper-and hypomethylation correlate with age, are selected and weighted in a linear model. The result is an equation, whereby chronological age can be estimated based on the percentage methylation at these key CpG sites in a given sample.
Epigenetic clocks have become increasingly diverse, with each predictor capturing different aspects of ageing. The expanding repertoire of clocks enable the study of ageing and rejuvenating approaches quantitatively. This review aims to give an overview of the growing toolbox of eAge clocks to inform which approach might best be suited to a scientific question.

| EPI G ENE TI C CLO CK S PRED I C T B I OLOG I C AL AG E
Epigenetic clocks have proven themselves to be accurate at predicting chronological age (chAge), which is commonly referred to as DNAm age or epigenetic age (eAge). When epigenetic clocks first emerged, a fundamental question arose; if eAge deviates significantly from chAge, is this difference due to inaccuracies of the clock itself, or caused by biological factors (e.g. genetics, disease status and environment)? In other words, are the clocks able to predict biological age? This difference between eAge and chAge is referred to as age acceleration and can be calculated as the mean absolute deviation (MAD) or median absolute deviation between eAge and chAge (Horvath, 2013), or as the residual from the linear regression between eAge and chAge . The reported error of epigenetic clocks can depend on how the test/validation data set was curated. For example, if a pooled data set were split into 90% samples for training and 10% for validation, then the clock would perform better on the validation data set than when applied to a completely independent/external data set. In the various tables surmising epigenetic clocks throughout this review, we will clarify how each clock is validated.
Vitamin D-sufficient individuals have a lower eAge acceleration and longer leukocyte telomere length (LTL) Vetter et al., 2020). Smoking has been associated with an increase in eAge in airway cells and lung tissue (4.9 and 4.3 years, respectively) (Wu et al., 2019), and smoking during pregnancy might have an effect on eAge in offspring (Simpkin et al., 2016). The number of studies associating eAge acceleration with diseases, phenotypes and environmental interventions that appear to affect ageing, emphasises eAge as a candidate metric for biological age .
However, some studies have shown no correlation between eAge acceleration and certain diseases or environmental factors, such as type II diabetes (Grant et al., 2017;Horvath, Gurven, et al., 2016), heroin use (Kozlenkov et al., 2017) or depression (Starnawska et al., 2019). Why eAge acceleration tracks with certain age-related disorders and not others are still not well understood. Most epigenetic clocks described in the following section used Illumina DNAm arraybased technology and are summarised in Table 1.

| Early epigenetic age predictors
The first epigenetic clocks incorporated relatively few CpG sites and samples in their training data sets, in comparison with later clocks.
Bocklandt et al., for example created a clock from 68 samples (34 twin pairs) that predicts age in saliva with an average accuracy of 5.2 years (Bocklandt et al., 2011). Koch and Wagner used five CpG sites and predicted age in multiple cell types, but with lower accuracy than the Bocklandt clock (MAD = 11 years) (Koch & Wagner, 2011). The same laboratory produced a six CpG clock that could track passage number in fibroblast cell cultures, regardless of original donor age (Koch et al., 2012). After these initial studies, epigenetic clocks grew in complexity in terms of number of samples, tissues and CpGs implemented.

| Multi-tissue age predictors
The first multi-tissue age predictor (referred to as the Horvath or Pan-Tissue clock) utilised 353 CpGs and has a mean error of 3.6 years, which at that time was unprecedented for any biomarker/age predictor (Horvath, 2013 For example, brain tissue appears to age slower relative to other tissues in the body, according to the Horvath clock (Horvath, 2013;Horvath, Mah, et al., 2015).
The association between age acceleration and health/disease status was first shown with the Horvath clock in obesity (Horvath et al., 2014) and has since become an established tool to assess biological age (Horvath et al., 2014;Horvath, Garagnani, et al., 2015;Horvath, Langfelder, et al., 2016;Maierhofer et al., 2017;Martin-Herranz et al., 2019). The Horvath clock has shown some limitations with particular tissues and ageassociated disease conditions. One of the most severe premature ageing syndromes, Hutchinson-Gilford Progeria Syndrome (HGPS), did not exhibit age acceleration according to the Horvath clock (Horvath, 2013). Children with multifocal developmental dysfunctions (syndrome X), who appear to age slower, do not decelerate in eAge (Walker et al., 2015). However, being rare genetic disorders, both studies were limited in the number of individuals tested.
The Horvath clock does not work reliably on cultured cells, particularly fibroblasts (Horvath, 2013;Horvath et al., , 2019. Replicative senescence in primary fibroblasts is a widely used model system in cellular ageing (Chandra & Kirschner, 2016;Hayflick, 1965;Hayflick & Moorhead, 1961  . This clock, known as the skin and blood (S&B) clock, is able to predict both in vivo and in vitro tissues accurately . The S&B clock also detected a modest, but significant age acceleration in HGPS samples .
The Zhang clock, while primarily trained to work on blood, is able to predict the ages of breast, liver, adipose and muscle tissue as accurately as the Horvath clock (Zhang, Vallerga, et al., 2019). This clock also outperformed both the Horvath and Hannum clocks in predicting blood age. It is set apart by the size of its training data with over 13,000 samples. Note: Epigenetic clocks created using a low number of CpGs (typically under 10), usually from preselected CpGs/regions known to have high age correlation. Error is based on mean absolute deviation (MAD).

| Tissue-specific age predictors
CpG sites that can predict age in blood, 99 of which were adapted for a separate clock that works on the Illumina 450 K array (Weidner et al., 2014). It was demonstrated that three of the 102 CpGs alone (selected by recursive feature elimination) can predict age in arrays and pyrosequenced samples (Lin et al., 2016;Weidner et al., 2014).
A minimal approach, such as this that uses as few CpGs as possible, only by targeting specific tissues, but also specific age groups.

| Minimised CpG clocks
Most of the epigenetic clocks mentioned above rely on Illumina Infinium arrays. The price of these arrays limits the applicability of eAge technology in drug discovery. Drug-discovery pipelines require the multiplexing of thousands of samples, but not necessarily the accuracy of the arrays. Here, we will discuss clocks relying on fewer CpGs (minimised clocks), which have the potential to be upscaled or run at a lower sample cost. The forensics field has developed multiple minimised clocks using strong age-associated CpGs (such as ELOVL2 and FHL2 (Garagnani et al., 2012;Bacalini et al., 2017)) and is designed for common tissues found at crime scenes, such as blood, saliva, buccal swabs and semen (Table 2).
Minimised clocks use a variety of technologies such as the Qiagen platform for pyrosequencing (referred to as pyrosequencing from here), which is more cost-effective for profiling the methylation of select CpGs. The Weidner 3 CpG clock (see "Tissue-Specific Age Predictors") for example, can predict age in blood samples using pyrosequencing (Weidner et al., 2014), but over-predicts age in saliva (a common source of DNA at crime scenes) by 14.6 years on average (Eipel et al., 2016). When adapted for saliva by adding two additional buccal-specific age-associated CpGs, eAge prediction was improved (Eipel et al., 2016).
Pyrosequencing had its own limitations (e.g. multiplexing; allowing a large number of samples and CpGs to be pooled and sequenced in a single run); however, new approaches increasing multiplex capabilities in pyrosequencing are emerging (Fleckhaus & Schneider, 2020). Another assay, termed SNaPshot, can multiplex 10 CpG sites (Thermo Fisher, 2020) and is used for many minimised epigenetic clocks.
The use of minimised clocks in forensics is just developing and for most clocks, cross-validation is missing (Cho et al., 2017). with SNaP-shot assays as tissue-specific age predictors, but also as a multi-tissue age predictor for common forensic tissues (blood, saliva and buccal swab) (Jung et al., 2019). Three of the CpG sites (near ELOVL2, FHL2 and C1orf132) have also proven sufficient to predict age efficiently (Dias, Cordeiro, Pereira, et al., 2020). These studies have demonstrated the versatility and accuracy predictors based on a few select CpGs can have and might be good candidates to increase the scale of eAge prediction.

| COMP OS ITE EPI G ENE TI C CLO CK S A S PRED I C TOR S OF MORB ID IT Y AND MORTALIT Y
Epigenetic clocks have proven capable of estimating not only chAge, but also time-to-death. Marioni et al. first showed that the higher difference between eAge and chAge, the greater the risk of all-cause mortality (mortality independent of health status, known genetic factors, and lifestyle factors) (Marioni, Shah, McRae, Chen et al., 2015). This finding was further validated in other studies (Chen et al., 2016;Christiansen et al., 2016). Positive age acceleration was also shown to predict cause-specific mortality in cancer and cardiovascular disease . These mortality associations were found using clocks that were not designed to directly predict mortality. Various composite approaches have been developed, whereby CpGs that correlate with metrics of physiological or cellular ageing (e.g. cholesterol or protein abundance) are used to construct a clock to predict age (Table 3). These clocks were built with the potential of capturing more of age-relevant biology than clocks trained on chAge alone.
The first composite biomarker age predictor (created independent of DNAm data) was based on 23 years of mortality data (Levine, 2013). This predictor incorporated ten biomarkers (e.g. C-reactive protein, glycated haemoglobin, systolic blood pressure, total cholesterol) that significantly correlated with age (Levine, 2013 proteins previously associated with mortality (Ignjatovic et al., 2011;Ridker et al., 2003), as surrogates for physiological risk factors (Lu, Quach, et al., 2019). The age acceleration of GrimAge was not only found to be associated with age-related conditions and lifestyle factors, but outperformed previous attempts at predicting timeto-death, time-to-coronary heart disease and time-to-cancer (Lu, Quach, et al., 2019). A significant association has been shown between GrimAge acceleration and lifelong trauma, but not childhood trauma (Katrinli et al., 2020), which is consistent with other studies showing Hannum (Wolf et al., 2016) and Horvath  clock age accelerations in post-traumatic stress disorder. GrimAge acceleration is also significantly associated with cortical atrophy (Katrinli et al., 2020), shorter pregnancy periods and lower birthweight .
Zhang et al. created a mortality-specific predictor, where they performed an epigenome-wide association study (EWAS) on a cohort with up to 14 years follow-up data. 58 CpGs were found that correlate with all-cause mortality, from which a predictor was constructed using only ten of the CpGs (Zhang, Wilson, et al., 2017). 48 of the CpGs identified had been associated with smoking, alcohol consumption, diabetes and cancer, some of which were also found in previous EWAS studies (Al Muftah et al., 2016;Chambers et al., 2015;Gao et al., 2015;Nilsson et al., 2014;Teschendorff et al., 2015;Travers et al., 2013;Zhang, Wilson, et al., 2017).
A DNAm telomere length (DNAmTL) estimator was created by Lu, Seeboth, et al. (2019), where leukocyte telomere length (LTL) was regressed against blood methylation data. This resulted in 140 LTL-associated CpGs forming the DNAmTL estimator (Lu, Seeboth, et al., 2019). Not only does DNAmTL predict LTL accurately, but it also demonstrates stronger predictive power of lifespan, time-tocoronary heart disease, time-to-congestive heart failure and smoking history compared to normal LTL. DNAmTL also showed associations with disease incidence for type 2 diabetes and ischaemic heart disease, respectively. Hence, composite clocks can use DNAm to predict non-DNAm traits, which in turn can be used as additional variables to accurately predict biological age, disease status and mortality.

| UNDERLYING MECHANIS M OF THE EPI G ENE TI C CLO CK
The Horvath clock is the most widely used clock for its accuracy, versatility and the accumulated knowledge we have of its behaviour from previous studies. It also gained the attention of the scientific community due to the fact that age can be predicted in multiple tissues using a relatively small number of CpGs (compared to the rest of the epigenome; Horvath, 2013). The fact that such a clock can be constructed provokes the question, is there a functional significance that correlates these CpGs with age in multiple tissues? If ageing is a phenomena that we are "programmed" to undergo, then are these CpGs an integral part of that machinery?
To understand the nature of eAge/epigenetic clocks, we must understand the aspects of physiological ageing they capture, the CpGs that constitute these clocks, and any causative relationships with ageing.

| Inaccuracies and tick-rate of epigenetic clocks
As with the Horvath clock, most clocks that followed after were also built on penalised linear regression models.  (Horvath, Pirazzini, et al., 2015). Similar findings were also found in analysis of cerebellum tissue from supercentenarians (Horvath, Mah, et al., 2015). The interpretation has been that the younger age predicted for centenarians reflects survival bias, where the lower biological age enabled the centenarians to live long.
However, with the clock possibly underpredicting older age systematically, this assumption might need to be reexamined. Alternatively, this discrepancy might be the result of a regression to the mean effect, where very high values (eAges) are underestimated by regression models.
While negative age acceleration (eAge predicted lower than chAge) was highest in the cerebellum, this underestimation was also observed in other tissues (including blood) from multiple data sets (El Khoury et al., 2019;Marioni et al., 2019;Martin-Herranz et al., 2019). It was also found that when accounting for age as a cofactor, the correlation between age acceleration and amyloid plaque load in brain tissue is attenuated (El Khoury et al., 2019), which is inconsistent with previous findings . It is possible that 5-hydroxymethyl cytosine (5hmC, an epigenetic modification more prevalent in brain tissue and indistinguishable from 5mC after bisulphite conversion) could cause age prediction offset in brain tissue (El (2019). These alterations in predictive accuracy of the clock in older individuals could be due to intrinsic changes in the rate of biological ageing during certain time points. The rate of change, or "tick" rate, was explored earlier in the Horvath clock study (Horvath, 2013). By looking at the weighted averages of the 353 CpGs compared with chAge, the tick rate was exponential between 0 and 20 years old, after which it continued linearly. As such, the Horvath clock applies a logarithmic transformation to ages <20 years, while the linear model is unaltered for ages >20 years (Horvath, 2013;Snir et al., 2019).
The study suggested that a higher organismal growth and cell division rate at early age might explain the initial acceleration in ageing (Horvath, 2013). A later study found a faster eAge tick rate during puberty in girls (Binder et al., 2018). However, no decrease in the tick rate of older subjects was observed (Horvath, 2013), which could be due to a lack of older individuals in the training data set used to construct the Horvath clock (El Khoury et al., 2019). Differences in tick rate could also be sex-specific. The Horvath, Hannum, and Zbiec´-Piekarska clocks show slightly faster ageing in men than women (Bergsma & Rogaeva, 2020).
A recent study found that simple multiple linear regression outperforms more involved machine learning techniques (Lau & Fung, 2020 Whether EPM or other non-linear models predict eAge in centenarians more accurately has not been determined.  . It would be intuitive to assume that eAge reflects other known aspects of ageing such as senescence, since an increase in senescence cells is considered a hallmark of ageing López-Otín et al., 2013). However, this has not been shown; instead, both replicative and damage-induced senescence do not correlate with increased eAge in vitro Lowe et al., 2016). It is possible that the accumulation of senescent cells in tissues with age remains proportionally low and that it is the effect on surrounding cells that is registered in eAge. Human telomerase reverse transcriptase (hTERT) expressing cells continue to epigenetically age despite never being able to enter replicative senescence . Leukocyte telomere length (LTL) erosion is one of the first biological phenomena that showed potential as biomarkers of ageing (Frenck et al., 1998;Harley et al., 1990;Hastie et al., 1990;Lindsey et al., 1991) and could be a physiological sign of ageing that correlates with eAge. However, like cellular proliferation and senescence, multiple studies have shown that eAge has no association with telomere length (Cypris et al., 2020;Horvath et al., 2019;Kabacik et al., 2018;Lowe et al., 2016;Marioni et al., 2016).

|
A plausible alternative is that eAge is governed by cellular dif-  (Huh et al., 2016), suggesting that direct reprogramming had no effect on eAge.
It has been hypothesised that eAge-related changes are reflected in intracellular alterations and changes in cell composition in a subset of cells termed "clock cells" . eAge might therefore capture the loss of somatic cells in some tissues  or the loss of stem cells, which do decline during ageing (Hernando-Herraez et al., 2019). A caveat is that eAge can be captured in neuronal cells, which are terminally differentiated cells and lack a stem cell pool (Horvath, 2013;Horvath, Mah, et al., 2015;. It also possible that eAge measures aspects of age-related epigenetic drift or deregulation . Demethylation can occur in either a passive manner (e.g. via inhibition of DNMT1 during cell replication; Wolffe et al., 1999;Mayer et al., 2000), or actively via methyl-CpG binding domain protein 4 (MBD4; Hendrich et al., 1999) or TET enzymes (Ichiyama et al., 2015;Jin et al., 2014). However, there is little evidence to suggest that active processes, such as TET, directly demethylate with age and affect eAge prediction (Wallace, 2014;Yu et al., 2020;Zhang et al., 2016). In addition, eAge can be measured in nonproliferating tissues (Horvath, 2013;Horvath et al., 2019;Yang et al., 2016) meaning passive demethyation is an unlikely mechanism. It is possible that actively dividing tissues accumulate somatic mutations in DNA methylation machinery during ageing, resulting in epigenetic drift observed as aberrant eAge prediction (Robertson et al., 2019).
The precise aspects of physiological ageing that eAge captures remain to be discovered, but further investigations into genes associated with eAge/clock CpGs and associations with other ageing biomarkers may disclose clues to the true nature of eAge.

| Causality of clock CpGs in ageing
DNAm became apparent as a potential biomarker of ageing with the discovery of strong age-associated CpGs, such as those in the CpG islands of ELOVL2, FHL2 and PENK1 (Bacalini et al., 2017;Garagnani et al., 2012). ELOVL2 is a strong biomarker for ageing in multiple tissues in both human and mouse (Bacalini et al., 2017;Chen et al., 2020;Garagnani et al., 2012;Hannum et al., 2013;Slieker et al., 2018). The CpGs neighbouring ELOVL2 strongly hypermethylate with age (Garagnani et al., 2012) and have been used in multiple forensic clocks (see "Minimised CpG Clocks"). ELOVL2 is an enzyme involved in elongation of long-chain polyunsaturated fatty acids, and also in the production of docosahexaenoic acid (DHA). DHA is the main polyunsaturated fatty acid in the retina and brain, and is necessary for healthy retinal function. Chen et al. showed that the Elovl2 promoter is more highly methylated in the retina of aged mice and that demethylation of this site recovers age-related decline in visual function via increased expression of Elovl2 . This is one of few studies to test a causal link of age-associated CpGs with phenotypic ageing.

| Transcriptional associations with eAge
One approach to functionally annotating CpGs is to analyse gene expression changes that correlate with the methylation of ageassociated CpGs. In the Horvath clock, the 193 CpGs (out of 353) that hypermethylate with age are more likely to be located in poised (bivalent) promoters. The 160 (out of 353) CpGs that hypomethylate with age are more likely to be in either weak promoters or strong enhancer regions (Horvath, 2013). However, linking the activity of age-related CpGs with specific gene expression has proven difficult Jung & Pfeifer, 2015;Yin et al., 2017;Zheng et al., 2016). The most likely reason is that many age-associated

| Genetic variants associated with eAge
A genome-wide association study (GWAS) is a method that could reveal genes that regulate eAge by finding genetic polymorphisms that correlate with eAge. A GWAS of cerebellum tissue found variants near an mTOR complex 2 gene (MLST8) and in an RNA-helicase gene (DHX57) that are associated with age acceleration. Many genes associated with cerebellar age acceleration also had overlap with neurodegenerative conditions such as Alzheimer's disease (Lu et al., 2016). Another GWAS revealed that one of the loci associated with intrinsic eAge acceleration (IEAA, which adjusts for both chAge and blood cell counts; Horvath, Gurven, et al., 2016;Quach et al., 2017) co-locates with hTERT . Variants of hTERT were found that associated with both IEAA and longer telomeres. Moreover, it was showed in vitro that higher hTERT expression (which is normally associated with cellular longevity) appears to cause a linear increases of eAge. By comparison, control cells passaged with no hTERT experienced an initial increase in eAge after 33 days in culture that eventually plateaued. These findings further enforce that eAge is not governed by cell division, replicative senescence or telomere length per se (Cypris et al., 2020;Horvath et al., 2019;Kabacik et al., 2018;Lowe et al., 2016;Marioni et al., 2016), since short telomeres are indicative of high proliferation and triggers replicative senescence. This paradoxical result could explain previous observations where during embryonic development and early postnatal life, the rate of epigenetic ageing is more rapid (Hiyama & Hiyama, 2007;Lu et al., 2018;Simpkin et al., 2016Simpkin et al., , 2017. These are periods of fast organismal growth coupled with high hTERT expression and cell division, which in turn would result in a higher eAge predicition.
Another approach to identify eAge-associated genetic vari-

| NON -H UMAN EPI G ENE TI C AG E PREDIC TOR S
Since the advent of DNAm age prediction for humans, age predictors have been created for other species; mice (Table 4), rats (Horvath, Singh, et al., 2020;Levine et al., 2020), dogs (Thompson et al., 2017;Wang et al., 2020), wolves (Thompson et al., 2017), humpback whales (Polanowski et al., 2014), chimpanzees (Guevara et al., 2020;Ito et al., 2018), marmosets (Horvath, Zoller, Haghani, Lu, et al., 2020), naked mole rats (Lowe et al., 2020), sea bass (Anastasiadi & Piferrer, 2020) and zebrafish (Mayne et al., 2020) (see Table 5 for a list of non-human/mouse epigenetic clocks). In 2017, three mouse epigenetic clocks were developed primarily using reduced representation bisulphite sequencing (RRBS) data. Wang et al. 2017 used 148 CpGs from liver tissue (using both RRBS and whole genome bisulphite data, WGBS) and found a moderate conservation of age-related CpGs between human and mouse. Their clock also showed an age reduction for calorie restriction, rapamycin and Prop1 df/df dwarfism (which results in lifespan extension up to 1.5 fold) (Brown-Borg et al., 1996;Cole et al., 2017;Wang et al., 2017). Petkovich et al. built a mouse epigenetic clock using 90 CpGs from blood and detected that calorie restriction reduces epigenetic age according to their clock (Petkovich et al., 2017). The first mouse multi-tissue age predictor was constructed based on 329 unique CpGs with a median absolute error of 3.33 weeks, mainly trained on young-and middle-aged mice (0.2-9.5 months) (Stubbs et al., 2017). A recent multi-tissue age predictor in mouse has been developed by Meer et al. that uses 435 CpGs, and predicts age across a wide age range (1-35 months) (Meer et al., 2018). It operates on multiple tissues including blood, liver, brain and heart (Meer et al., 2018). Thompson et al. created four mouse RRBS clocks to compare statistical methods and found the most accurate clock resulting from elastic net regression (Thompson et al., 2018).
The Wang, Stubbs and Petkovich mouse clocks mentioned here show little overlap in CpGs used (Field et al., 2018). This is probably due to the variability of RRBS data, where the regional genome coverage differs between protocols and enzymes used, rather than different statistical methods applied (Field et al., 2018;Thompson et al., 2018 Indeed, other studies have created a similar custom array to accurately predict age in model organisms. Currently available as a preprint, the Horvath laboratory has published an epigenetic clock that works on both rats and humans (Horvath, Singh, et al., 2020). This was created using a custom Illumina methylation array called the HorvathMammalMethylChip40, made up of 36,000 CpGs conserved among 50 mammalian species Horvath, Singh, et al., 2020). The MAE for human and rat data was 0.03, and a correlation of 0.95. Three single tissue clocks were also created for rat liver, brain and blood, as well as a multitissue clock combining all three tissues (Horvath, Singh, et al., 2020). Another preprint has been released of a sheep epigenetic clock, using the same array, with a median error of 5.1 months (~3.5-4.2% of expected sheep lifespan). The study reported that castrated sheep had a higher age acceleration than age-matched controls, and a dual human and sheep clock was constructed with an additional 1,848 human samples (Sugrue et al., 2020). A rat clock has also been developed using 134 RRBS whole blood samples . Elastic net selected 68 CpGs, and had a correlation of r = 0.9 in their test data set. It appears to work in mice, where it predicted reduced age acceleration after calorie restriction.
Many age-associated CpGs are conserved between different species (Horvath, 2013;Horvath, Singh, et al., 2020;Wang et al., 2017Wang et al., , 2020, meaning the development of pan-species clocks is plausible. For example, an epigenetic clock has been created using 394 CpGs from modules of developmental genes with conserved, age-related methylation changes, between mouse, human and dogs . Recent preprints have shown various universal pan-tissue epigenetic clocks that predict age across 9 tissue types from 128 different mammalian species , and models that predict maximum lifespan, gestation time and sexual maturity . The CpGs used to construct the clocks were also associated with genes that are enriched during mammalian development . These clocks further enforce the idea that ageing is a conserved evolutionary process intertwined with mammalian development.

| CON CLUS ION
eAge prediction is a powerful approach that has revolutionised experimental gerontology. As the number and diversity of epigenetic clocks increases, so too does our understanding of biological age. Multiple studies have shown that it is possible to build accurate minimised clocks using only a few highly age-associated CpGs (Cho et al., 2017;Daunay et al., 2019;Dias, Cordeiro, Pereira, et al., 2020;Jung et al., 2019;Weidner et al., 2014;Zbiec´-Piekarska et al., 2015). Many of these clocks use CpGs nearby ELOVL2 and FHL2, and work in saliva and blood (Cho et al., 2017;Dias, Cordeiro, Pereira, et al., 2020;Jung et al., 2019;Zbiec´-Piekarska et al., 2015).
They have yet to be tested in other scenarios, such as clinical applications. On the other hand, clocks might be more robust when utilising a large number of CpGs (Boroni et al., 2020;Zhang, Kamath ,et al., 2019).
The approach used by Horvath to develop epigenetic clocks has spawned not only an abundance of similar DNAm age predictors, but also other novel approaches, such as transcriptional (Bryois et al., 2017;Peters et al., 2015), proteomic (Lehallier et al., 2019;Tanaka et al., 2018) and cellular biophysical/biomolecular (Phillip et al., 2017) clocks. Indeed, DNAm can be regressed with health co-factors such as smoking and alcohol consumption to produce predictors of complex traits and mortality (McCartney et al., 2018). While DNAm is one of the most accurate and versatile biomarkers for ageing and disease, our understanding of it is still developing. Perhaps looking at DNAm in combination with other non-DNAm based biomarkers will broaden our understanding and predictive power of biological ageing and mortality. Composite clocks such as PhenoAge and GrimAge are first steps in that direction. Transcription clocks may reveal regulators of biological ageing, for example, if key ageing genes are found to be linked with eAge either by correlating with age acceleration or directly with methylation changes of key clock CpGs.
To us, the key areas to emerge will be in a) understanding the different aspects of ageing captured by distinct clocks and b) testing causality of DNAm in age acceleration through interventional epigenetics and other approaches. As epigenetic clocks become more sophisticated and commonplace, caution must be considered when inferring the biological significance of age acceleration.
Research must continue regarding the nature of eAge and the aspects of ageing captured.

ACK N OWLED G EM ENTS
D.J.S. is funded by the Medical Research Council (Doctoral Training Programme in Precision Medicine) and T.C. is a Chancellor's Fellow at the University of Edinburgh. We thank Riccardo Marioni for critical reading of the manuscript.

CO N FLI C T O F I NTE R E S T
None declared.