Min Woo Sun
Aging Clocks: Ticking Away from Birthday Candles to Biological Age
March 24, 2021
All humans shall eventually succumb to the inevitable march of time. At least, that is how it is today. Aging is defined as the “progressive decline in functional integrity and homeostasis, culminating in death” . It is associated with increased risk for numerous diseases that often do not manifest until old age from cancer to Alzheimer’s disease. Furthermore, the biology of aging has not been well understood due to its complexity. Is it possible to proactively halt the decline in physiological functions and damage wrought by aging?
From the philosopher’s stone to Qin Shi Huang’s quest for the elixir of life, humanity has sought for the panacea of longevity and even immortality throughout history. While such aspirations seemed futile and hopeless, the advent of high throughput sequencing technologies amassing tens of thousands of molecular data and the rise of computational power met with clever algorithms has allowed researchers to investigate aging scientifically with data-driven approaches rather than blindly following unsubstantiated hearsays and snake oil. Interest in aging has been growing among the general public and in the industry as well. A mainstream podcast, the Joe Rogan Experience recently featured Dr. Aubrey De Grey and Dr. David Sinclair exploring various topics and progress in aging research [2, 3]. The Rejuvenation Roadmap curated by LifeSpan.io enlists numerous academic groups and companies pursuing longevity therapies with many undergoing clinical trials . Aging research is no longer out of reach, but an active field of interest. Despite this growing enthusiasm and endeavor, aging remains to be largely an enigma. The definition of aging is often not agreed upon and the molecular underpinnings of aging is still unclear, shrouded by the complex interaction between genetic and environmental factors. From epigenetic alterations to mitochondrial dysfunction, aging has been characterized by various hallmarks, just showing how complicated the process of aging can be, involving countless pathways and systems . Aging has also been implicated in cognitive decline and reduced diversity in the human gut microbiome just to name a few associations [6, 7]. The manifestation of aging also presents heterogeneously depending on biological sex . To comprehend such a complex biological phenomenon, I argue that one of the top priorities in current aging research is accurately quantifying and predicting age. As Galileo Galilei once said, we must “measure what is measurable, and make measurable what is not so.” For both assessing the efficacy of interventional treatments for longevity as well as understanding the etiology of aging, methods to quantify and predict biological age is of utmost importance to aging research. In fact, many leading researchers in the aging field have been focusing more and more attention towards novel ways to measure and predict age.
“How old are you?” Our response to this seemingly simple question is most often our chronological age based on calendar years–counting the years that have passed since our birthday. The notion of age is closely tied to time; however, chronological age does not necessarily represent the actual biological age of our body. In fact, quantifying biological age is a lot more involved than simply counting the number of candles on a birthday cake. With the rise of machine learning and artificial intelligence, statisticians and computer scientists have developed novel methods that can help find insights in complex biological data that often pose the challenge of high dimensionality (eg. common SNPs for genome-wide association studies) as well as heterogeneity in varying data modalities (e.g. multiomic data). As much as there is bench science in the lab that allows for experiments and generation of data, bioinformatics and quantitative approaches are necessary to process and make meaningful conclusions from the data. Many of the methods employed to predict chronological and biological age relies on this very balance between molecular data produced from labs and clever models involving regressions or machine learning algorithms. Here, in this paper we will explore the various aging clocks that aim to predict chronological and biological age and show that developing novel ways to measure age is allowing researchers to achieve deeper insights into the biology of aging.
Epigenetic clocks: DNAm Age and DNAm PhenoAge [8, 9]
One of the most popular aging clocks is based on the epigenome, specifically methylation. Methylation occurs when a methyl group is added to a cytosine nucleotide base in the genome. As methylation does not permanently alter the nucleotide base, it is considered an epigenetic mechanism. Methylation has been shown to play an integral role in gene regulation and has been used for research in cancer to functional genomics, even identifying the tissue from which cell-free DNA found in the peripheral blood originated from . In the context of aging, there has been mounting evidence that indicates epigenetic factors can be used to predict age. For instance, specific CpG sites (cytosine followed by guanine nucleotide base) exhibit abnormal methylation–hypo- or hypermethylation–patterns that vary with age . Based on these associations, Dr. Horvath collected publicly available Illumina DNA methylation array data (n=7844) in order to construct an age predictor generating DNAm age . In order to fit a model and avoid overfitting, i.e. the model cannot generalize well to data it has not been shown during the training process, Dr. Horvath split the datasets into a training and validation set. Each of the CpG sites that were found in both Illumina methylation platforms served as a feature (also known as independent variable or predictor), leading to a large feature space–a common problem in biological data where the number of features are greater than the number of samples, i.e. p > n. In such supervised learning scenarios, it is common to implement a model that penalizes complexity of the model. Dr. Horvath fitted an elastic net model which is a linear regression that utilizes both the lasso (L1 norm) and ridge (L2 norm) balancing the two penalties using weights. A 10-fold cross validation was used in order to choose the best shrinkage factor lambda, which determines how much penalization will have influence in the objective function. Often used as a framework for selecting model parameters, k-fold cross validation splits the training data set into k equally sized folds and utilizes each fold for validation and the remaining folds for training the model. These procedures allowed the model to narrow down to 353 CpG sites from a total of 21,369 initial CpG sites. The weighted average of these selected features are considered as the epigenetic clock that is used to fire predictions, dubbed DNAm age.
The DNAm age generated from this model performed remarkably well across multiple tissues, exhibiting high correlation with chronological age even in the test dataset that was not used during the model’s training process. Furthermore, an Ingenuity Pathway Analysis of the genes harboring the 353 CpG sites that were identified through the model were enriched for pathways in “cell death/survival, cellular growth/proliferation, organismal/tissue development, and cancer” showing association with aging . Pathway analysis generally identifies shared pathways for a set of given genes or proteins. These are not random sites throughout the genome, but specific regions that may be involved in the process of aging. According to the paper, some of the useful properties of DNAm age include the fact that the prediction values are near zero for pluripotent stem cells and can be used to measure age acceleration regarded to be highly heritable. The multi-tissue age predictor generating DNAm age can be freely accessed in https://dnamage.genetics.ucla.edu/.
More recently, the work on DNAm has been further expanded to be more comprehensive and show higher performance. While DNAm age used chronological age as a proxy for biological age, DNAm PhenoAge utilizes phenotypic age based on multiple clinical biomarkers . This allows for the prediction of morbidity and mortality outcome rather than just chronological age. In order to determine the relevant clinical measurements to use as phenotypic age, Levine et al. got 42 clinical biomarkers such as white blood cell count and mean cell volume and fit a regression model predicting aging-related mortality like Alzheimer’s disease). The biomarkers are from clinical data from National Health and Nutrition Examination Survey (NHANES). NHANES aims to measure the health and nutritional status of people in the United States and combines both interviews and physical examinations .
This approach selected 10 biomarkers including chronological age to be used for phenotypic age predictor, which will also serve as the target variable (the response or dependent variable in supervised learning) for the final DNAm PhenoAge model. The phenotype age is then regressed on the methylation data set, similar to DNAm age model using linear regression with elastic net penalization. This led to a selection of 513 CpG sites, many of these lying in CpG islands (stretches of repeating cytosine followed by guanine). 41 of these CpGs overlap with the 353 CpG sites identified using the DNAm age predictor suggesting similarity, but also novel targets. Note that the CpGs that were chosen and used in the DNAm predictor were optimized to be predictive of chronological age, whereas CpGs for DNAm PhenoAge predictor optimized for the selected biomarkers characterizing phenotype age. The additional CpGs are conducive to predicting phenotype based on biological age rather than just chronological age. This suggests that DNAm PhenoAge can be more flexibly used in predicting not just chronological, but biological age than DNAm.
Proteomic aging clock 
While epigenetic clocks based on methylation have shown promising results, there are other types of aging clocks that aim to predict chronological or phenotypic age using different data modalities. Protein based aging clocks have also shown to be highly predictive of chronological age. In their recent works, Lehallier et. al identified 529 common plasma proteins that exhibit expression level changes associated with aging by analyzing a proteomic dataset through linear regression coefficients (including the all subset method to reduce feature space) . Further looking into common plasma aging proteins through literature search showed close links to aging, including experiments demonstrating proteins that can affect the lifespan of animals in animal models. All these associations suggest that building an aging clock utilizing proteomic data could lead to a highly performative age predictor.
Similar to DNAm and DNAm PhenoAge, Lehallier et. al fit linear models including ridge or lasso penalties on a set of training data and tested on a validation set not exposed during the training procedure. As lasso induces sparsity in the model and can send coefficients to zero, the lasso penalty model allows for subset selection of proteins. Model performance was measured using Pearson correlation between predicted age and true age as well as median absolute error. The most predictive clock among the numerous models attempted (different subsets of proteins considered for regression) was the lasso with correlation 0f 0.98 for training set and 0.96 for the validation set and median absolute error of 1.84 years and 2.44 years respectively. The model was further tested on two independent plasma proteomic datasets, which also exhibited high performance for the model (correlation 0.9, 0.91).
The proteomic clock was also able to show that inactive patients have a larger predicted age (predicted from the proteomic clock, i.e. previously described lasso model) than their chronological age. This suggests that the proteomic clock can characterize human health to some degree. One downside of a protein based aging clock is that the cost of getting the protein data is high, but by reducing the number of proteins considered in the aging clock through model regularization, this particular proteomic aging clock reduces the fiscal burden of using protein data.
FRIGHT and AFRAID clock 
Rather than building a model using molecular data like methylation or common plasma proteins, Schultz et al. employed frailty indices to build aging clocks. Frailty indices are based on various biometrics that quantify health deficits and are usually non-invasive. In humans, frailty indices have been shown to have high predictive power for mortality and morbidity, even better compared to DNA methylation based aging clocks . Using frailty indices measured longitudinally from mice, Schultz et al. built two novel aging clocks, FRIGHT (Frailty Inferred Geriatric Health Timeline) for predicting chronological age and AFRAID (Analysis of Frailty and Death) for predicting life expectancy. Examples of frailty indices for mice used in this study include tail stiffening, breathing rate/depth, and hearing loss.
In order to build a predictor for chronological age, multivariate linear regression models (including elastic net), random forest, and Klemera-Doubal methods (KDMs)–all fitted on frailty index items–were considered along with the frailty index score as a single predictor. These supervised learning approaches on frailty index items performed better than just using the frailty index score on its own. The random forest model had the lowest error rate and strong correlation with chronological age compared to other methods leading to the FRIGHT clock. Random forest–a method that fits many decision trees to make predictions–are capable of capturing nonlinear relationships as well as interactions between the features and the response that is not possible using linear approaches.
While FRIGHT is useful for predicting chronological age, it is not suited for predicting mortality. In contrast, AFRAID aims to predict life expectancy. Similar to FRIGHT, linear regression with simple least squares and elastic net, and random forest were considered. Not surprisingly, elastic net and random forest performed much better than a simple linear regression with least squares method as they incorporate regularization and thus reduce overfitting. The random forest model was also chosen as the final model for AFRAID.
Both FRIGHT and AFRAID clocks have been empirically shown their potential for serving as biomarkers to assess the efficacy of treatments. For a study that treated 21 mice with angiotensin-converting enzyme (ACE) inhibition enalapril, FRIGHT age for the treated mice was a month younger than the control mice that did not receive the enalapril; however, AFRAID did not predict that the treated mice would survive longer. This is consistent with the analysis that enalapril can improve health, but not increase the maximum lifespan for mice. Furthermore, both FRIGHT and AFRAID are based on fragility indices that are non-invasive and easier to measure compared to molecular data like methylation or protein, which would require repeated phlebotomy for longitudinal observation.
From telomere length to state-of-the-art machine learning based age predictors, more and more novel techniques for quantifying and predicting biological age are being developed based on new methods or data modality. As we have seen, many of these approaches follow a supervised learning framework, relying on fitting a model–even as simple as good old linear regression–on often high dimensional biological data to predict age. Studying complex biological processes like aging requires the synergistic endeavor of biologists and experts in quantitative and computational fields. Interdisciplinary effort is key to aging research and pursuing creative solutions to such a challenging problem.
A clear trend in aging clocks reviewed in this paper is the application of linear models to fit a predictor to the data. In particular, linear regression with penalty accounting for model complexity was commonly deployed for many of the aging clocks. Choice of linear model allows for the interpretation of the variables and their respective marginal influence over the response variable. This is generally more difficult to retrieve from blackbox methods like eXtreme Gradient Boosting (XGB) or neural networks. However, these linear approaches are limited to capturing parametric (making assumptions about the data, e.g. assuming normality) and linear relationships and will be biased for nonlinear relationships.
Similar to the application of the random forest model in FRIGHT and AFRAID clock, I believe future aging clocks will incorporate more complex methods like deep learning for increased predictive performance at the cost of interpretation. Such approaches could be more appropriate for biological data with complex relationships and interactions between the features and the target variable that we may be unaware of or too difficult to measure. Another avenue for improvement could be the utilization of mutliomic data, i.e. integrating various omics data sets such as genome, transcriptome, and proteome. Multiomic data fusion could allow the aging clocks to be exposed to the molecular phenomenon that can only be observed in a particular omics layer or complex interactions that can be seen comprehensively when multiple data are combined. For instance, epigenetic changes will be missed if the model is only shown genomic data. However, acquiring such datasets can get costly despite the rapidly decreasing cost of sequencing molecular data.
Developing accurate ways to quantify and predict both chronological and biological age are crucial to aging research. In particular, highly predictive aging clocks can be useful for evaluating the efficacy of treatments and intervention for increasing human longevity and healthspan, which has long been a bottleneck. Simply waiting for decades to observe the effects of these treatments in humans will not lead to the necessary funding and regulatory approvals. As such, without a way to measure the effect of these procedures in humans, it is difficult to make meaningful progress. Fortunately, aging clocks reviewed in this paper are examples of diverse solutions developed by leading scientists and labs around the world to solve this very problem. Future improvements in aging clocks will lead to further progress in aging research and longevity treatments.
- Singh, P. P., Demmitt, B. A., Nath, R. D., & Brunet, A. (2019). The genetics of aging: A vertebrate perspective. Cell, 177(1), 200-220. doi:10.1016/j.cell.2019.02.038
- #1432: Solving the Aging Problem – Aubrey de Grey (2020, February). In The Joe Rogan Experience.
- #1349: David Sinclair (2019, January). In The Joe Rogan Experience.
- The rejuvenation roadmap. (n.d.). Retrieved March 21, 2021, from https://www.lifespan.io/road-maps/the-rejuvenation-roadmap/
- López-Otín, C., Blasco, M. A., Partridge, L., Serrano, M., & Kroemer, G. (2013). The hallmarks of aging. Cell, 153(6), 1194-1217. doi:10.1016/j.cell.2013.05.039
- Murman, D. (2015). The impact of age on cognition. Seminars in Hearing, 36(03), 111-121. doi:10.1055/s-0035-1555115
- Popkes, M., & Valenzano, D. R. (2020). Microbiota–host interactions shape Ageing dynamics. Philosophical Transactions of the Royal Society B: Biological Sciences, 375(1808), 20190596. doi:10.1098/rstb.2019.0596
- Fraga, M. F., & Esteller, M. (2007). Epigenetics and aging: The targets and the marks. Trends in Genetics, 23(8), 413-418. doi:10.1016/j.tig.2007.05.008
- Levine, M. E., Lu, A. T., Quach, A., Chen, B. H., Assimes, T. L., Bandinelli, S., . . . Horvath, S. (2018). An epigenetic biomarker of aging for lifespan and healthspan. doi:10.1101/276162
- Moss, J., Magenheim, J., Neiman, D., Zemmour, H., Loyfer, N., Korach, A., . . . Dor, Y. (2018). Comprehensive human cell-type methylation atlas reveals origins of circulating cell-free dna in health and disease. Nature Communications, 9(1). doi:10.1038/s41467-018-07466-6
- Fraga, M. F., & Esteller, M. (2007). Epigenetics and aging: The targets and the marks. Trends in Genetics, 23(8), 413-418. doi:10.1016/j.tig.2007.05.008
- NHANES – about the national health and Nutrition Examination Survey. (2017, September 15). Retrieved March 24, 2021, from https://www.cdc.gov/nchs/nhanes/about_nhanes.htm
- Lehallier, B., Shokhirev, M. N., Wyss‐Coray, T., & Johnson, A. A. (2020). Data mining of human plasma proteins generates a multitude of highly predictive aging clocks that reflect different aspects of aging. Aging Cell, 19(11). doi:10.1111/acel.13256
- Johnson, A. A., Shokhirev, M. N., Wyss-Coray, T., & Lehallier, B. (2020). Systematic review and analysis of human proteomics aging studies unveils a novel proteomic aging clock and identifies key processes that change with age. Ageing Research Reviews, 60, 101070. doi:10.1016/j.arr.2020.101070
- Schultz, M. B., Kane, A. E., Mitchell, S. J., MacArthur, M. R., Warner, E., Vogel, D. S., . . . Sinclair, D. A. (2020). Age and life expectancy clocks based on machine learning analysis of mouse frailty. Nature Communications, 11(1). doi:10.1038/s41467-020-18446-0
- Kim, S., Myers, L., Wyckoff, J., Cherry, K. E., & Jazwinski, S. M. (2017). The frailty index outperforms dna methylation age and its derivatives as an indicator of biological age. GeroScience, 39(1), 83-92. doi:10.1007/s11357-017-9960-3