Skip to main content
Epigenomics logoLink to Epigenomics
. 2024 Nov 25;17(1):49–57. doi: 10.1080/17501911.2024.2432854

Insights to aging prediction with AI based epigenetic clocks

Joshua J Levy a,b,c,d,e,✉, Alos B Diallo f, Marietta K Saldias Montivero f, Sameer Gabbita g, Lucas A Salas e,h, Brock C Christensen e,i
PMCID: PMC11703013  PMID: 39584810

ABSTRACT

Over the past century, human lifespan has increased remarkably, yet the inevitability of aging persists. The disparity between biological age, which reflects pathological deterioration and disease, and chronological age, indicative of normal aging, has driven prior research focused on identifying mechanisms that could inform interventions to reverse excessive age-related deterioration and reduce morbidity and mortality. DNA methylation has emerged as an important predictor of age, leading to the development of epigenetic clocks that quantify the extent of pathological deterioration beyond what is typically expected for a given age. Machine learning technologies offer promising avenues to enhance our understanding of the biological mechanisms governing aging by further elucidating the gap between biological and chronological ages. This perspective article examines current algorithmic approaches to epigenetic clocks, explores the use of machine learning for age estimation from DNA methylation, and discusses how refining the interpretation of ML methods and tailoring their inferences for specific patient populations and cell types can amplify the utility of these technologies in age prediction. By harnessing insights from machine learning, we are well-positioned to effectively adapt, customize and personalize interventions aimed at aging.

KEYWORDS: DNA methylation, epigenetics, aging, clocks, artificial intelligence, classification and regression trees, deep learning, explainability

1. Introduction

Aging is an inevitable process marked by the gradual deterioration of physiological functions, manifesting as increased cellular senescence, genomic instability, metabolic alterations, and other biomolecular changes. Normal aging involves gradual changes in physical and cognitive functions, impacting some daily activities. However, subjects must be able to adapt with a minimal impact on quality of life. On the other hand, pathological aging involves degenerative diseases in different organ systems, including neurodegeneration and dementia, cardiovascular conditions, and joint/muscle impairments, among others. Frailty syndromes are the cumulative sum of multiple systems declining, leading to higher morbidity and mortality of aging individuals [1]. Understanding the mechanisms of normal and pathological aging is crucial for deciphering the developmental processes from youth to old age, addressing the health challenges and adaptations required with advanced years, and potentially early interventions for those with decreased organ reserve to prevent frailty syndromes and diseases linked with pathological aging. The concept of aging often invokes a mix of fear and fascination; it is inexorably linked to the specter of disease and the certainty of mortality. Despite this, or perhaps because of it, society harbors a deep obsession with aging and certain biases involving ageism. It is viewed as our ultimate adversary and a formidable challenge to overcome.

While human lifespan has significantly increased over the past century, this extension has not been accompanied by a commensurate reduction in chronic diseases, leading to extended periods of morbidity [2,3]. This reality underscores the need to lengthen life and enhance quality of life. From the public health perspective, we are now balancing the pressures of adapting to an aged society while preventing suffering and reducing the pressure of higher economic costs for the individual and society. Achieving this requires a nuanced understanding of the highly variable biological processes that govern aging – distinguishing between what is expected versus what are pathological changes. By understanding our biological age, which can diverge from our chronological age (Figure 1(a)), we can better monitor our health and tailor interventions to optimize our well-being.

Figure 1.

Figure 1.

Epigenetic clocks and future advances. (a) Biological vs. Chronological age to illustrate the concept of age acceleration: on the left, a datapoint has been added for a specific patient representing where biological age surpasses chronological age suggests aging processes that is faster than expected. Conversely, in the middle, a biological age less than chronological age for a specific patient indicates the influence of biological processes and lifestyle factors that contribute to a slower aging process. On the right, the goal of many interventions is to slow down the process of aging for a specific patient as they age, represented in this graphic by the patient (data point) moving from an accelerated aging state to a decelerated over time. (b) Age prediction methods: graphical representation of classical methods such as EWAS (1 CpG to predict age as indicated using green arrow), penalized multivariable regression (e.g., LASSO, elastic Net; penalized approach models many CpGs simultaneously while selecting a subset that are the most predictive of age), and advanced machine learning techniques like random forest and deep learning (graphic at bottom left). These advanced methods can further minimize the residual between biological and chronological ages with the potential to provide insights into the underlying mechanisms. This is represented graphically by plotting the same patient twice as a two data points – the traditional regression approaches (blue arrow/line) estimate a larger residual between biological and chronological age, whereas the machine learning approaches estimate a smaller residual (red arrow/line). (c) Explainable AI techniques: depicts the use of explainable AI to organize CpGs into genes and pathways within deep learning models, enhancing the interpretability of epigenetic data. DNAm beta values representing the proportion of methylated alleles at individual CpG sites are aggregated and transformed based on related genes to form gene summaries of DNAm. Similarly, gene-level DNAm information is summed and aggregated on the pathway level. Green arrows indicate this mapping between CpGs to genes to pathways. (d) CpG and Cell Type Interactions: discusses the integration of CpG sites with cell type data estimated through deconvolution approaches. Beta values at specific CpG sites for one patient are represented using the gray squares in the upper left of the panel. Cell type abundance for the same patient is represented through the circles in the upper right of the panel, with different colors for each cell type. This panel explores how CART models can elucidate molecular mechanisms behind age acceleration specific to certain cells or patient subpopulations when derived interactions are applied to generalized linear modeling approaches. The decision tree selects specific cell types and CpGs from which decision rules are made, further binning patients based on these decisions. The green arrows indicate the logical flow between the splits. If a decision split based on a cell-type follows a CpG decision split and vice versa, there exists a potential statistical interaction between a CpG and cell-type. These interactions are tested through statistical modeling on the right, which demonstrates how the association between a specific CpG’s methylation status and age can change depending on the cell-type (i.e., different slope for blue versus yellow cell-type).

The last decade and a half have witnessed significant advances in algorithmic methods designed to determine our biological age [4,5]. Among the various analytes available for determining biological age, epigenetics – particularly through DNA methylation (DNAm) – holds the most promise. The prediction of age using epigenetic/DNAm data is referred to as an epigenetic clock [6]. DNAm is the addition of a methyl group primarily to cytosine in the context of a cytosine – phosphate – guanine dinucleotide (CpG). It is essential to cellular lineage specification during development, exerting effects on gene expression without altering the underlying DNA sequence. A typical assay for DNAm is a form of modified genotyping using microarray-based technology, such as the Illumina HumanMethylationEPIC array, which measures the proportion of methylated alleles at > 900,000 CpGs genome-wide [7]. In addition, reduced sequencing costs are enabling even more comprehensive measures of DNAm in larger sample sizes more practical with reduced representation bisulfite sequencing (RRBS), whole genome bisulfite sequencing (WGBS) [8–10], and new panel-based approaches like the Human Methylome Panel from TWIST that measures 4 M CpG sites [11]. With increased accuracy, the direct calling of cytosine modifications with long-read sequencing like Oxford Nanopore Technology, and through paired-end sequencing technologies developed by Biomodal, which offers “6-base” sequencing, is also possible. Aging is associated with DNAm alterations regulated by enzymes such as DNA methyltransferases (DNMTs), responsible for adding and maintaining these marks throughout our lifespan [12,13]. Consequently, specific sites may show dynamic increases or decreases in DNAm over time. Many of these modifications are closely tied to metabolic pathways, highlighting the essential role of metabolic alterations and diet in their relationship with inflammation, as observed in centenarian populations. Other epigenetic biomolecules, such as histone modifications and non-coding RNAs (ncRNAs), are also implicated in the regulation of gene expression and are associated with aging through mechanisms like post-transcriptional suppression or activation [14]. For example, long non-coding RNAs (lncRNAs) have been linked to cellular differentiation, replicative senescence, and age-related cognitive decline [15,16]. Similarly, small ncRNAs (e.g., miRNAs, circRNAs) have been associated with gene silencing and the regulation of cellular senescence, particularly in cardiac and vascular fibroblast and endothelial cell lineages [17,18]. Histone modifications, including ubiquitylation, methylation, and acetylation, are hypothesized to influence the activation and silencing of longevity-associated genes, with global histone loss and changes in acetylation patterns observed during aging [19]. However, much of this research has been conducted in model organisms, with limited human studies, and thus a detailed discussion of these epigenetic modifications is beyond the scope of this perspective article.

2. Epigenetic clocks and the promise of machine learning

Epigenetic clocks are algorithmic approaches that summarize changes across numerous age-associated CpG sites to predict chronological age [20]. However, the prediction is referred to as biological age, which can differ from chronological age. This difference, the model’s residual, reflects deviations from expected aging patterns and is used to identify age acceleration or deceleration. For example, increased age acceleration may indicate exposure to environmental stressors, suboptimal dietary habits, or lifestyle choices that negatively affect one’s physiological age [21]. These factors include both modifiable behaviors, such as diet and physical activity, and less controllable intrinsic mechanisms. Conversely, a decreased rate of age acceleration often mirrors a movement toward positive health behaviors and conditions such as a nutritious diet, regular exercise, a healthy lifestyle, and a supportive built environment. Since DNA methylation represents a reversible somatic alteration, research has demonstrated the possibility of inducing epigenetic modifications that can effectively “slow down or turn back” the biological clock [22]. These changes can be triggered by pharmacological treatments or modifications in lifestyle, offering promising avenues for intervention in the aging process.

Epigenetic clocks originally employed generalized linear modeling techniques [20,23,24]. This approach typically involves identifying CpG sites where DNAm changes linearly with age. These sites are then incorporated into an elastic net penalized model, which selects a smaller, non-redundant set of CpGs for age estimation. The age prediction is made through a linear combination of these CpGs, providing an intuitive interpretation of age acceleration based on hypo- or hypermethylation relative to a general population. A key example of this interpretation includes the methylation alterations in the ELOVL2 (ELOVL fatty acid elongase 2) promoter, which affect the biosynthesis of very long-chain fatty acids in cell membranes [25]. However, this straightforward explanation is based on simplified assumptions. In reality, the aging process is far more complex. The rate of DNAm changes can vary across different developmental stages – clocks such as GrimAge and PhenoAge incorporate nonlinear components to account for these variations [26–29]. Epigenetic clocks have been developed using DNAm measures in biospecimens containing mixtures of cell types (often peripheral blood), where variation in cell type proportions with age [12,13], and cell-type-specific DNAm patterns [30,31] can both impart effects. These complexities add layers that are challenging to conceptualize but are crucial for explaining the discrepancies in residual values between chronological age and biological age from epigenetic clocks.

Machine learning (ML) technologies represent a set of computational heuristics that excel in capturing complex associations. These include methods like Classification and Regression Trees (CART), which split data into branches based on feature values, and deep learning, a subset of machine learning that uses layered artificial neural networks to analyze data at various levels of abstraction [32,33]. These technologies introduce model-building procedures that can identify, from nearly an infinite possible set of combinations of nonlinear transformations and interactions, the set that optimally approximates the relationship between DNAm and aging. A high level technical summary of these methodologies can be found in the Supplementary Material (see section “Additional Technical Overview of Machine Learning Methods”). By reducing the residual between chronological and biological age, ML approaches have the potential to offer deeper insights into the aging process (Figure 1(b)). However, achieving a complete, intuitive understanding of the resulting age acceleration measures remains elusive, partly due to the challenges involved in interpreting these advanced models.

3. Explainable epigenetic clocks

Explainable AI (XAI) techniques can enhance our understanding of age acceleration derived from complex models. XAI facilitates the identification and interpretation of CpG sites and their interactions, shedding light on their broader implications, such as their influence on biological pathways involved in aging and age acceleration [34]. These methods can be implemented post hoc, exemplified by tools like Shapley Additive Explanations (SHAP), which construct local additive models to elucidate interactions between CpG sites for each sample [35]. Alternatively, XAI can be seamlessly integrated during the model-building phase. Recent advancements in deep learning have concentrated on hierarchizing CpG sites into modules or capsules, which can be structured according to pathways at various scales (Figure 1(c)) [36,37].

The interpretability of a model largely depends on the chosen machine learning method, as each offers unique advantages in its ability to predict and decipher the aging process. For instance, CART models are particularly effective at capturing statistical interactions but produce stepwise predictions, which may artificially increase the residual. Conversely, deep learning and methods like elastic net provide a smoother output and better capture of the continuous nature of biological aging. Bayesian Additive Regression Trees (BART) have the potential to smooth age predictions from CART approaches by treating each decision tree as a random variable [38]. BART enhances accuracy by averaging predictions across multiple realizations of each tree and across the entire ensemble of trees within a given posterior sample. Averaging across the predictive posterior distribution can yield less biased and more predictive age estimates. However, scaling these approaches to high-dimensional feature spaces remains a challenge.

4. Multimodal approaches

There is considerable potential in leveraging CART within more complex frameworks, such as leveraging statistical interactions learned by these models to inform elastic net generalized linear mixed models (GLMMs) and facilitate more intuitive interpretations without overwhelming computational demands [39].

There is considerable potential to leverage ML approaches to combine DNAm data with other data types (e.g., demographic data) to enhance the accuracy of epigenetic clocks as well as their personalization. For instance, CART models can uncover nuanced CpG-demographic interactions, which can then be analyzed using routine statistical approaches like GLMMs. Using standard statistical methods, this reanalysis can better appreciate the statistical significance and effect size of potential effect modifiers, enabling the examination of distinct aging pathways or variations in age acceleration across different patient subgroups. Not only is aging a critical risk factor for cancer, but beyond chronological age, epigenetic age acceleration has been shown to be an independent risk factor for cancer in the prospective setting [40]. Such differential patterns of aging highlight the potential for personalized aging research, enabling targeted interventions based on specific demographic characteristics. Precision cancer prevention approaches that track and intervene in biological aging in adults offer promise to reduce cancer burden and could be deployed in the primary care setting.

Cellular changes play a crucial role in mediating and modifying our understanding of molecular pathways involved in aging. DNAm is particularly effective for estimating age and cellular proportions through advances in reference-based cell-type deconvolution approaches. Recent work by Ze Zhang et al. has highlighted how cellular abundance, particularly in naïve and memory cell type fractions, is a source of residual confounding in epigenetic clocks [41]. Future iterations of epigenetic clocks should distinguish between cell-type specific changes and those intrinsic to aging processes [42]. Developing new epigenetic clocks capable of estimating age acceleration, dependent on and independent of cellular proportions, could provide valuable insights into the critical mechanisms of age acceleration [39]. In addition, distinguishing cell type dependency of aging effects will improve the practical opportunities for precision interventions on biological aging and healthy immune aging. This dependency analysis is another area where CART models can be employed to identify differential methylation of age by cell type (Figure 1(d)) [6]. Though largely unexplored, incorporating methylation cytometry into ML-based epigenetic clocks can further enhance our understanding of how cellular composition changes affect aging.

CART approaches are limited in their ability to integrate more complex data modalities. In contrast, deep learning approaches offer greater flexibility, allowing them to combine more seamlessly with complex data types such as gigapixel histopathological slides, single-cell assays, and spatial genomics assays [43].

There is potential utility in using age acceleration measures derived from epigenetic clocks to understand concurrent changes across various assays. These approaches can be integrated with new devices, wearables, and imaging technologies and correlated with other assays to comprehend their effects better [44]. For example, DNAm has been used to predict measures related to the visual aging of human skin [45]. Similar clocks are being developed for retinal images and brain MRI features which have been associated with genomic indicators of aging and cognitive impairment respectively [46]. By stripping away these components, additional aspects of age acceleration that are not tied to these well-known factors can be revealed. For instance, facial wrinkling often reflects collagen depletion; removing the effect of this confounder can identify less obvious underlying factors [14,47–49]. Alternatively, DNAm-derived age acceleration can be predicted from these other data types to uncover additional aging features and can serve as lower-cost surrogate assessments [26].

5. Opportunities to enhance the predictiveness and adaptability of epigenetic clocks

Looking forward, numerous opportunities exist to enhance epigenetic clocks’ predictiveness. For instance, little work has been done to understand how epigenetic clocks developed using specific cohorts or conditions can be adapted to other domains. Enhancing predictiveness and statistical precision would permit the development of clocks using smaller, less-powered cohorts.

Deep learning models trained to predict age from one cohort can be repurposed or fine-tuned in another cohort through a process known as transfer learning [33–35,50]. Parameters from fully trained models can serve as initial parameters for training on a new cohort, reducing the amount of data needed, especially if the new cohort shares characteristics similar to those of the previous one. Bayesian approaches also offer this flexibility, with posterior parameter distributions from one cohort serving as priors for another [51]. This approach may be particularly useful for developing pan-tissue clocks [52]. While clocks have been trained across various tissue types simultaneously, it is crucial to determine the relevance of these approaches. Overemphasizing commonalities between tissues may overlook tissue-specific aging effects that warrant consideration [28,53]. Further work is needed to apply ML methods to study tissue-specific aging, especially given the growing importance of epigenetic clocks in defining the hallmarks of precancer and cancer risk [54].

The quality control and relevance of CpG sites may vary between cohorts, making prediction more challenging due to the differing availability of these sites. This may necessitate adopting or replacing missing information with adequate surrogate sites or using approaches with greater flexibility in handling missing data [55–57]. Future work on implementing epigenetic clocks should consider the minimal set of CpGs needed to enhance prediction, allowing the development of targeted DNAm sequencing or custom microarray chip assays that can be deployed at population scale for routine biological age and cell type proportions assessment. Given additional sites from sequencing-based technologies, the robustness of these clocks remains uncertain.

Additionally, bisulfite conversion used in both array-based and sequencing technologies does not distinguish between 5-methylcytosine (5mC) and 5-hydroxymethylcytosine (5hmC), where 5hmC signal is read as 5mC. This is an area where long-read and nanopore sequencing could potentially be beneficial in the future. Far less is known about 5hmC and aging. It should be noted that 5hmC is more prominent in stem cell populations and generally absent in terminally differentiated cells. In addition, reductions of 5hmC have been linked to dementia/Alzheimer’s disease [58,59]. Together, these findings suggest potential benefits for incorporating 5hmC into future epigenetic clocks.

6. Potential clinical implications

Opportunities to understand and intervene on biological aging have the potential to provide humankind with collective leaps in quality life-years with reduced morbidity and mortality. At a minimum, even regular physical activities at very small doses to prevent problems with balance and falling in normal aging – these changes can soon be registered through application of next generation epigenetic clocks.

There are many potential clinical implications for this technology of note, including:

Identifying individuals at-risk for age-related conditions: A meta-analysis found that accelerated clocks are linked to all-cause mortality and predict future cancer, diabetes, COPD, and cardiovascular diseases [60,61]. Additional studies have associated acceleration with metabolic syndromes and BMI, and deceleration with physical activity, as well as links to Alzheimer’s disease [21,62]. However, variability in associations across studies may be due to the limited set of CpGs used, which may not be the most biologically informative, and the influence of the training populations [63]. Nonlinear relationships between age acceleration and disease incidence, influenced by chronological age, may also play a role. Machine learning models may be more adept at integrating information from a larger set of CpGs, reflecting more complex biological pathways, potentially improving their consistency in predicting future age-related diseases.

Monitoring intervention effectiveness. Age acceleration derived through these epigenetic clocks have been used as baseline and endpoint markers to assess the effectiveness of various interventions aimed at slowing down or reversing biological aging. Most research to date has focused on interventions such as diet and exercise, but ongoing trials are expanding to include assessments of interventions targeting circadian rhythm calibration, smoking cessation, and inflammation reversal [16,22,64–69]. Other promising areas under investigation include the efficacy of topical treatments, geroprotective drugs, and novel interventions such as small molecules, senolytics, and cellular reprogramming.

Personalize treatment options. With the shift toward epigenetic clocks that target specific cell types, demographics, diverse data types, surrogate markers, and biological pathways, the potential to identify precise biological drivers of accelerated aging is increasing [70–74]. This enables a more tailored approach to treatments designed to slow or reverse aging in these specific areas. For example, recent studies utilizing single-cell methylation profiling in mice have further enhanced our understanding of aging dynamics at the cellular level over time [73].

7. Ethical considerations

Questions also persist regarding the integration of these emerging monitoring technologies, therapeutics, and treatment options into existing reimbursement frameworks. These discussions are pivotal for guiding startup ventures focused on these areas. Additionally, recent FDA guidelines that restrict the use of Laboratory Developed Tests (LDTs) could significantly influence the adoption of machine learning technologies for aging monitoring. Understanding these regulatory impacts is essential for navigating the future landscape of aging intervention technologies, though it should be noted that there are already commercial ventures that have proven quite successful in providing biological age measures and nutraceuticals [75].

These challenges also raise ethical concerns about access to these technologies and the potential for discrimination based on predicted biological age. For instance, epigenetic clocks are being considered as underwriting tools for life insurance companies and medical practices, but this could have far-reaching implications. Individuals with access to these tests might be motivated to take out life insurance policies based on their results, raising questions about the fairness of such practices. Furthermore, age acceleration reflects not only individual risk factors but also broader social and structural influences beyond personal control [76,77]. This raises concerns about the impact on vulnerable groups, as these tools may inadvertently reinforce existing inequalities. Careful consideration should be given to the implications of requiring consent from all individuals, particularly in ensuring that vulnerable populations are not disproportionately affected or excluded [78].

Enhancing transparency and explainability may improve fairness in the use of these predictive models. As discussed in the XAI section, methods like SHAP can provide valuable insights into how predictions are made, helping to distinguish between risk factors that are deliberate and those beyond individual control. However, SHAP is associative rather than causal, necessitating careful interpretation [79–81]. Educating stakeholders, particularly those involved in the development and implementation of these models in decision-making, could be beneficial in promoting a responsible understanding of their limitations. Furthermore, it has been suggested that clearly communicating the uncertainty inherent in model predictions could encourage critical evaluation and disrupt rote interpretation of findings [82–86]. This is particularly important as models may be less trained on certain demographics, leading to greater uncertainty in predictions, even if not necessarily reduced accuracy.

Recent literature on the social implications of epigenetic clocks has proposed the adoption of a human rights-based framework to assess the fairness of these models [77]. Such a framework would involve evaluating whether there is potential for discriminatory intent, disproportionate impacts on specific groups, and whether fundamental rights, such as informed consent, are upheld. Additionally, this framework would aim to ensure that biological age metrics accurately reflect intentional lifestyle choices rather than external factors beyond individual control, though distinguishing between these influences remains a significant challenge. As machine learning models continue to evolve, they may prioritize different factors based on their design and development [87]. This variability underscores the need for guidelines to ensure that epigenetic clocks are developed and implemented in a manner that is both equitable and effective, while adequately representing diverse populations.

8. Future perspective

This perspective article has explored the continued development and evolving role of AI-based epigenetic clocks and their potential to enhance our understanding of aging by identifying both reducible and irreducible factors. These tools hold significant promise for identifying individuals at risk of age-related diseases, monitoring the effectiveness of interventions, and personalizing treatment options. However, further research is needed to fully understand how the integration of more advanced AI technologies will enhance their clinical application and impact. Future application of machine learning technologies will provide opportunities to further disentangle the complexities of aging in ways that can be tailored to specific demographics or used to distinguish cellular components of accelerated aging. Future developments in this area may also involve the adoption of algorithmic strategies such as Bayesian analysis and transfer learning to improve model predictiveness while reducing the amount of data required, along with integrative multimodal analyses that combine DNAm with other complex data types. Additionally, challenges remain in interpretation and deployment of these sophisticated models in practice without introducing new regulatory and ethical considerations. Therefore, future efforts in this field may benefit from strategies that prioritize the collection and representation of data in ways that minimize the potential for misuse, ensuring responsible and ethical applications. As we venture further into these uncharted territories, only time will reveal the broader implications of our quest to slow down and turn back the biological clock.

Supplementary Material

Supplemental Material

Acknowledgments

We would like to thank Zarif Azher, California Institute of Technology, for his insightful discussion.

ChatGPT-4 was used to help improve the grammar, clarity and flow of the manuscript and was not used beyond this capacity.

Funding Statement

Joshua Levy is supported by Department of Defense grant PR220927, and NIH awards P20GM130454, P20GM104416, R24GM141194, R01CA277810. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Article highlights

  • Introduction: Understanding the biological mechanisms that govern aging and deviations from expected aging is crucial for addressing adaptations needed for a longer, healthier life with reduced morbidity. DNA methylation, a reversible epigenetic mark strongly correlated with age, holds great potential in deciphering these mechanisms and understanding age-related disease vulnerabilities.

  • Epigenetic Clocks and the Promise of Machine Learning: Epigenetic clocks are algorithms designed to predict chronological age from DNA methylation data, which serves as a proxy for biological age. The residual between biological and chronological age represents age acceleration. Traditionally based on generalized linear models, the integration of machine learning has enabled these clocks to capture nuanced, nonlinear, and interaction effects, improving the understanding of discrepancies between predicted and actual age. Supplementary information has been introduced which provides nuanced comparisons between machine learning approaches for age prediction.

  • Explainable Epigenetic Clocks: AI-based epigenetic clocks often face challenges in interpretability. Explainable AI techniques can help address these issues by identifying biological pathways associated with aging. The interpretability of clocks depends on the modeling approach – whether it is deep learning or decision trees.

  • Multimodal Approaches: Integrating diverse data types into epigenetic clocks, such as cell-type abundance or cohort demographics, allows machine learning models to pinpoint cell-type or demographic-specific methylation patterns linked to aging. Additionally, incorporating other assays and imaging modalities can reveal new associations, helping to isolate novel biological pathways contributing to age acceleration.

  • Opportunities to Enhance the Predictiveness and Adaptability of Epigenetic Clocks: Approaches like transfer learning and Bayesian analysis offer opportunities to improve the predictiveness of epigenetic clocks while reducing the data needed by leveraging prior knowledge or information from other cohorts. Challenges remain in balancing the need to identify the minimal, nonredundant set of age-related CpGs for cost-efficient predictions while avoiding over-reliance on certain biological pathways.

  • Potential Clinical Implications: There is significant potential for AI-enhanced epigenetic clocks to use age acceleration measures to identify individuals at risk for age-related diseases, monitor the effectiveness of anti-aging interventions, and personalize treatment options.

  • Ethical Considerations: The introduction of epigenetic clocks into underwriting practices may raise ethical concerns, particularly if these models reflect factors beyond individual control. Enhancing the fairness and transparency of prediction models, while adopting a human rights framework that minimizes disproportionate impacts on vulnerable groups, will be crucial to ensuring equitable application.

  • Future Perspective: The role of AI in developing epigenetic clocks requires careful oversight to ensure ethical use and minimize biases, particularly regarding pathway prioritization of CpG site selection and cohort-specific associations. Future research may benefit from personalizing clocks for diverse populations, elucidating cell-type-specific pathways for tailored treatments, and evaluating the clinical robustness of age acceleration measures across machine learning algorithms. A focus on gathering diverse, high-quality data to improve the consistency and explanatory power of these clocks may prove more impactful than solely advancing algorithmic techniques.

Disclosure statement

The authors have no relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

No writing assistance was utilized in the production of this manuscript.

Author contributions

All authors contributed to the writing and editing of this manuscript.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/17501911.2024.2432854

References

Papers of special note have been highlighted as either of interest (•) or of considerable interest (••) to readers.

  • 1.Chen X, Mao G, Leng SX.. Frailty syndrome: an overview. Clin Interv Aging. 2014;9:433–441. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Crimmins EM. Lifespan and Healthspan: past, present, and promise. Gerontologist. 2015;55:901–911. doi: 10.1093/geront/gnv130 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Partridge L, Deelen J, Slagboom PE. Facing up to the global challenges of ageing. Nature. 2018;561:45–56. doi: 10.1038/s41586-018-0457-8 [DOI] [PubMed] [Google Scholar]
  • 4.Bafei SEC, Shen C. Biomarkers selection and mathematical modeling in biological age estimation. NPJ Aging. 2023;9:1–9. doi: 10.1038/s41514-023-00110-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Jia L, Zhang W, Chen X. Common methods of biological age estimation. Clin Interv Aging. 2017;12:759–772. doi: 10.2147/CIA.S134921 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Bell CG, Lowe R, Adams PD, et al. DNA methylation aging clocks: challenges and recommendations. Genome Biol. 2019;20:249. doi: 10.1186/s13059-019-1824-y [DOI] [PMC free article] [PubMed] [Google Scholar]; • Helpful review and commentary on DNA methylation based epigenetic clocks.
  • 7.Pidsley R, Zotenko E, Peters TJ, et al. Critical evaluation of the illumina MethylationEPIC BeadChip microarray for whole-genome DNA methylation profiling. Genome Biol. 2016;17:208. doi: 10.1186/s13059-016-1066-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Meissner A, Gnirke A, Bell GW, et al. Reduced representation bisulfite sequencing for comparative high-resolution DNA methylation analysis. Nucleic Acids Res. 2005;33:5868–5877. doi: 10.1093/nar/gki901 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Zhou L, Ng HK, Drautz-Moses DI, et al. Systematic evaluation of library preparation methods and sequencing platforms for high-throughput whole genome bisulfite sequencing. Sci Rep. 2019;9:10383. doi: 10.1038/s41598-019-46875-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Miura F, Enomoto Y, Dairiki R, et al. Amplification-free whole-genome bisulfite sequencing by post-bisulfite adaptor tagging. Nucleic Acids Res. 2012;40:e136–e136. doi: 10.1093/nar/gks454 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Olova N, Krueger F, Andrews S, et al. Comparison of whole-genome bisulfite sequencing library preparation strategies identifies sources of biases affecting DNA methylation data. Genome Biol. 2018;19:33. doi: 10.1186/s13059-018-1408-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Fraga MF, Ballestar E, Paz MF, et al. Epigenetic differences arise during the lifetime of monozygotic twins. Proc Natl Acad Sci. 2005;102:10604–10609. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Christensen BC, Houseman EA, Marsit CJ, et al. Aging and environmental exposures alter tissue-specific DNA methylation dependent upon CpG island context. PLOS Genet. 2009;5:e1000602. doi: 10.1371/journal.pgen.1000602 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Orioli D, Dellambra E. Epigenetic regulation of skin cells in natural aging and premature aging diseases. Cells. 2018;7:268. doi: 10.3390/cells7120268 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Pereira Fernandes D, Bitar M, Jacobs FMJ, et al. Long non-coding RNAs in neuronal aging. Noncoding RNA. 2018;4:12. doi: 10.3390/ncrna4020012 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Wang K, Liu H, Hu Q, et al. Epigenetic regulation of aging: implications for interventions of aging and diseases. Sig Transduct Target Ther. 2022;7:1–22. doi: 10.1038/s41392-022-01211-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Dhahbi JM. Circulating small noncoding RNAs as biomarkers of aging. Ageing Res Rev. 2014;17:86–98. doi: 10.1016/j.arr.2014.02.005 [DOI] [PubMed] [Google Scholar]
  • 18.Gupta SK, Piccoli MT, Thum T. Non-coding RNAs in cardiovascular ageing. Ageing Res Rev. 2014;17:79–85. doi: 10.1016/j.arr.2014.01.002 [DOI] [PubMed] [Google Scholar]
  • 19.Wang Y, Yuan Q, Xie L. Histone modifications in aging: the underlying mechanisms and implications. Curr STEM Cell Res Ther. 2018;13:125–135. doi: 10.2174/1574888X12666170817141921 [DOI] [PubMed] [Google Scholar]
  • 20.Binder AM, Horvath S. Epigenetic clocks. 2022. [cited 2024 Sep 2]. Available from: https://link.springer.com/chapter/10.1007/978-3-030-94475-9_11
  • 21.Oblak L, van der Zaag J, Higgins-Chen AT, et al. A systematic review of biological, social and environmental factors associated with epigenetic clock acceleration. Ageing Res Rev. 2021;69:101348. doi: 10.1016/j.arr.2021.101348 [DOI] [PubMed] [Google Scholar]
  • 22.Fahy GM, Brooke RT, Watson JP, et al. Reversal of epigenetic aging and immunosenescent trends in humans. Aging Cell. 2019;18:e13028. doi: 10.1111/acel.13028 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Simpson DJ, Chandra T. Epigenetic age prediction. Aging Cell. 2021;20:e13452. doi: 10.1111/acel.13452 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Horvath S, Raj K. DNA methylation-based biomarkers and the epigenetic clock theory of ageing. Nat Rev Genet. 2018;19:371–384. doi: 10.1038/s41576-018-0004-3 [DOI] [PubMed] [Google Scholar]
  • 25.Garagnani P, Bacalini MG, Pirazzini C, et al. Methylation of 2 gene as a new epigenetic marker of age. Aging Cell. 2012;11:1132–1134. doi: 10.1111/acel.12005 [DOI] [PubMed] [Google Scholar]
  • 26.Lu AT, Quach A, Wilson JG, et al. DNA methylation GrimAge strongly predicts lifespan and healthspan. Aging (Albany NY). 2019;11:303–327. doi: 10.18632/aging.101684 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Levine ME, Lu AT, Quach A, et al. An epigenetic biomarker of aging for lifespan and healthspan. Aging (Albany NY). 2018;10:573–591. doi: 10.18632/aging.101414 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Horvath S. DNA methylation age of human tissues and cell types. Genome Biol. 2013;14:R115. [DOI] [PMC free article] [PubMed] [Google Scholar]; •• Landmark work on the development of the first DNA methylation epigenetic clock.
  • 29.Lu AT, Fei Z, Haghani A, et al. Universal DNA methylation age across mammalian tissues. Nat Aging. 2023;3:1144–1166. doi: 10.1038/s43587-023-00462-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Houseman EA, Accomando WP, Koestler DC, et al. DNA methylation arrays as surrogate measures of cell mixture distribution. BMC Bioinformatics. 2012;13:86. doi: 10.1186/1471-2105-13-86 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Salas LA, Zhang Z, Koestler DC, et al. Enhanced cell deconvolution of peripheral blood using DNA methylation for high-resolution immune profiling. Nat Commun. 2022;13:761. doi: 10.1038/s41467-021-27864-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Loh W-Y. Classification and regression trees. Wiley Interdiscip Rev: Data Min Knowl Discov. 2011;1:14–23. doi: 10.1002/widm.8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521:436–444. doi: 10.1038/nature14539 [DOI] [PubMed] [Google Scholar]
  • 34.de Lima Camillo LP, Lapierre LR, Singh R. A pan-tissue DNA-methylation epigenetic clock based on deep learning. NPJ Aging. 2022;8:1–15.35927252 [Google Scholar]
  • 35.Levy JJ, Titus AJ, Petersen CL, et al. MethylNet: an automated and modular deep learning approach for DNA methylation analysis. BMC Bioinformatics. 2020;21:108. doi: 10.1186/s12859-020-3443-8 [DOI] [PMC free article] [PubMed] [Google Scholar]; • One of the first deep learning models developed for DNAm age prediction.
  • 36.Levy JJ, Chen Y, Azizgolshani N, et al. MethylSPWNet and MethylCapsNet: biologically motivated organization of DNAm neural networks, inspired by capsule networks. NPJ Syst Biol Appl. 2021;7:1–16. doi: 10.1038/s41540-021-00193-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Prosz A, Pipek O, Börcsök J, et al. Biologically informed deep learning for explainable epigenetic clocks. Sci Rep. 2024;14:1306. [DOI] [PMC free article] [PubMed] [Google Scholar]; • An example of explainable AI techniques includes the use of deep learning to identify age-relevant biological pathways by organizing CpGs according to their gene and pathway associations.
  • 38.Tan YV, Roy J. Bayesian additive regression trees and the general BART model. Stat Med. 2019;38:5048–5069. doi: 10.1002/sim.8347 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Levy JJ, O’Malley AJ. Don’t dismiss logistic regression: the case for sensible extraction of interactions in the era of machine learning. BMC Med Res Methodol. 2020;20:171. doi: 10.1186/s12874-020-01046-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.Kresovich JK, Xu Z, O’Brien KM, et al. Methylation-based biological age and breast cancer risk. J Natl Cancer Inst. 2019. doi: 10.1093/jnci/djz020 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41.Zhang Z, Reynolds SR, Stolrow HG, et al. Deciphering the role of immune cell composition in epigenetic age acceleration: insights from cell-type deconvolution applied to human blood epigenetic clocks. Aging Cell. 2024;23:e14071. doi: 10.1111/acel.14071 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 42.Buckley MT, Sun ED, George BM, et al. Cell-type-specific aging clocks to quantify aging and rejuvenation in neurogenic regions of the brain. Nat Aging. 2023;3:121–137. doi: 10.1038/s43587-022-00335-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43.Azher ZL, Suvarna A, Chen J-Q, et al. Assessment of emerging pretraining strategies in interpretable multimodal deep learning for cancer prognostication. BioData Min. 2023;16:23. doi: 10.1186/s13040-023-00338-w [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44.Petersen CL, Christensen BC, Batsis JA. Weight management intervention identifies association of decreased DNA methylation age with improved functional age measures in older adults with obesity. Clin Epigenet. 2021;13:46. doi: 10.1186/s13148-021-01031-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 45.Bienkowska A, Raddatz G, Söhle J, et al. Development of an epigenetic clock to predict visual age progression of human skin. Front Aging [Internet]. 2024. [cited 2024 Jul 19];4. Available from: https://www.frontiersin.org/journals/aging/articles/10.3389/fragi.2023.1258183/full [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46.Meng D, Zhang S, Huang Y, et al. Application of AI in biological age prediction. Curr Opin Struct Biol. 2024;85:102777. doi: 10.1016/j.sbi.2024.102777 [DOI] [PubMed] [Google Scholar]
  • 47.de Oliveira NFP, de Souza BF, de Castro Coêlho M. UV radiation and its relation to DNA methylation in epidermal cells: a review. Epigenomes. 2020;4:23. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 48.Holzscheck N, Söhle J, Schläger T, et al. Concomitant DNA methylation and transcriptome signatures define epidermal responses to acute solar UV radiation. Sci Rep. 2020;10:12918. doi: 10.1038/s41598-020-69683-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 49.Marioni RE, Belsky DW, Deary IJ, et al. Association of facial ageing with DNA methylation and epigenetic age predictions. Clin Epigenetics. 2018;10:140. doi: 10.1186/s13148-018-0572-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 50.Urban A, Sidorenko D, Zagirova D, et al. Precious1GPT: multimodal transformer-based transfer learning for aging clock development and feature importance analysis for aging and age-related disease target discovery. Aging (Albany NY). 2023;15:4649–4666. doi: 10.18632/aging.204788 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 51.McElreath R. Statistical rethinking: a bayesian course with examples in R and Stan. Boca Raton, FL, USA: CRC press; 2020. [Google Scholar]
  • 52.Handl L, Jalali A, Scherer M, et al. Weighted elastic net for unsupervised domain adaptation with application to age prediction from DNA methylation data. Bioinformatics. 2019;35:i154–i163. doi: 10.1093/bioinformatics/btz338 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 53.Boroni M, Zonari A, Reis de Oliveira C, et al. Highly accurate skin-specific methylome analysis algorithm as a platform to screen and validate therapeutics for healthy aging. Clin Epigenetics. 2020;12:105. doi: 10.1186/s13148-020-00899-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 54.Stangis MM, Chen Z, Min J, et al. The hallmarks of precancer. Cancer Discov. 2024;14:683–689. doi: 10.1158/2159-8290.CD-23-1550 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 55.Di Lena P, Sala C, Prodi A, et al. Missing value estimation methods for DNA methylation data. Bioinformatics. 2019;35:3786–3793. doi: 10.1093/bioinformatics/btz134 [DOI] [PubMed] [Google Scholar]
  • 56.Lena PD, Sala C, Prodi A, et al. Methylation data imputation performances under different representations and missingness patterns. BMC Bioinformatics. 2020;21:268. doi: 10.1186/s12859-020-03592-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 57.McEwen LM, Jones MJ, Lin DTS, et al. Systematic evaluation of DNA methylation age estimation with common preprocessing methods and the infinium MethylationEPIC BeadChip array. Clin Epigenetics. 2018;10:123. doi: 10.1186/s13148-018-0556-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 58.López V, Fernández AF, Fraga MF. The role of 5-hydroxymethylcytosine in development, aging and age-related diseases. Ageing Res Rev. 2017;37:28–38. doi: 10.1016/j.arr.2017.05.002 [DOI] [PubMed] [Google Scholar]
  • 59.Coppieters N, Dieriks BV, Lill C, et al. Global changes in DNA methylation and hydroxymethylation in Alzheimer’s disease human brain. Neurobiol Aging. 2014;35:1334–1344. doi: 10.1016/j.neurobiolaging.2013.11.031 [DOI] [PubMed] [Google Scholar]
  • 60.Salameh Y, Bejaoui Y, El Hajj N. DNA methylation biomarkers in aging and age-related diseases. Front Genet. 2020;11:480672. doi: 10.3389/fgene.2020.00171 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 61.Field AE, Robertson NA, Wang T, et al. DNA methylation clocks in aging: categories, causes, and consequences. Mol Cell. 2018;71:882–895. doi: 10.1016/j.molcel.2018.08.008 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 62.Fransquet PD, Wrigglesworth J, Woods RL, et al. The epigenetic clock as a predictor of disease and mortality risk: a systematic review and meta-analysis. Clin Epigenetics. 2019;11:62. doi: 10.1186/s13148-019-0656-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 63.McCrory C, Fiorito G, Hernandez B, et al. GrimAge outperforms other epigenetic clocks in the prediction of age-related clinical phenotypes and all-cause mortality. J Gerontology: Ser A. 2021;76:741–749. doi: 10.1093/gerona/glaa286 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 64.Fitzgerald KN, Hodges R, Hanes D, et al. Potential reversal of epigenetic age using a diet and lifestyle intervention: a pilot randomized clinical trial. Aging (Albany NY). 2021;13:9419–9432. doi: 10.18632/aging.202913 [DOI] [PMC free article] [PubMed] [Google Scholar]; •• Example of using epigenetic age acceleration as a measure to identify the effectiveness of various health interventions.
  • 65.Fiorito G, Caini S, Palli D, et al. DNA methylation-based biomarkers of aging were slowed down in a two-year diet and physical activity intervention trial: the DAMA study. Aging Cell. 2021;20:e13439. doi: 10.1111/acel.13439 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 66.Moqri M, Herzog C, Poganik JR, et al. Biomarkers of aging for the identification and evaluation of longevity interventions. Cell. 2023;186:3758–3775. doi: 10.1016/j.cell.2023.08.003 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 67.Yaskolka Meir A, Keller M, Hoffmann A, et al. The effect of polyphenols on DNA methylation-assessed biological age attenuation: the DIRECT PLUS randomized controlled trial. BMC Med. 2023;21:1–15. doi: 10.1186/s12916-023-03067-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 68.Kawamura T, Radak Z, Tabata H, et al. Associations between cardiorespiratory fitness and lifestyle-related factors with DNA methylation-based ageing clocks in older men: WASEDA’S health study. Aging Cell. 2024;23:e13960. doi: 10.1111/acel.13960 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 69.Sae-Lee C, Corsi S, Barrow TM, et al. Dietary intervention modifies DNA methylation age assessed by the epigenetic clock. Mol Nutr Food Res. 2018;62:1800092. doi: 10.1002/mnfr.201800092 [DOI] [PubMed] [Google Scholar]
  • 70.Mandelblatt JS, Ahles TA, Lippman ME, et al. Applying a life course biological age framework to improving the care of individuals with adult cancers: review and research recommendations. JAMA Oncol. 2021;7:1692–1699. doi: 10.1001/jamaoncol.2021.1160 [DOI] [PMC free article] [PubMed] [Google Scholar]; • Example application of biological age to guide disease management and personalize treatments.
  • 71.Vidaki A, Ballard D, Aliferi A, et al. DNA methylation-based forensic age prediction using artificial neural networks and next generation sequencing. Forensic Sci Int: Genet. 2017;28:225–236. doi: 10.1016/j.fsigen.2017.02.009 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 72.Griñán-Ferré C, Bellver-Sanchis A, Guerrero A, et al. Advancing personalized medicine in neurodegenerative diseases: the role of epigenetics and pharmacoepigenomics in pharmacotherapy. Pharmacol Res. 2024;205:107247. doi: 10.1016/j.phrs.2024.107247 [DOI] [PubMed] [Google Scholar]
  • 73.Trapp A, Kerepesi C, Gladyshev VN. Profiling epigenetic age in single cells. Nat Aging. 2021;1:1189–1201. [DOI] [PMC free article] [PubMed] [Google Scholar]; • Future epigenetic clocks should consider deepening our understanding of cell-type specific effects.
  • 74.Crimmins EM, Thyagarajan B, Levine ME, et al. Associations of age, sex, race/ethnicity, and education with 13 epigenetic clocks in a nationally representative US sample: the health and retirement study. J Gerontol Ser A. 2021;76:1117–1123. doi: 10.1093/gerona/glab016 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 75.Moqri M, Herzog C, Poganik JR, et al. Validation of biomarkers of aging. Nat Med. 2024;30:360–372. doi: 10.1038/s41591-023-02784-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 76.Lawrence KG, Kresovich JK, O’Brien KM, et al. Association of neighborhood deprivation with epigenetic aging using 4 clock metrics. JAMA Network Open. 2020;3:e2024329–e2024329. doi: 10.1001/jamanetworkopen.2020.24329 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 77.Dupras C, Beck S, Rothstein MA, et al. Potential (mis)use of epigenetic age estimators by private companies and public agencies: human rights law should provide ethical guidance. Environ Epigenet. 2019;5:dvz018. [Google Scholar]; • Introduces ethical framework guiding the further development and usage of epigenetic clocks amongst related concerns.
  • 78.Salas LA, Peres LC, Thayer ZM, et al. A transdisciplinary approach to understand the epigenetic basis of Race/Ethnicity health disparities. Epigenomics. 2021;13:1761–1770. doi: 10.2217/epi-2020-0080 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 79.Ma S, Tourani R. Predictive and causal implications of using Shapley value for Model interpretation. In: Proceedings of the 2020 KDD Workshop on Causal Discovery [Internet]. PMLR; 2020. p. 23–38. [cited 2022 May 21]. Available from: https://proceedings.mlr.press/v127/ma20a.html [Google Scholar]
  • 80.Kumar IE, Venkatasubramanian S, Scheidegger C, et al. Problems with shapley-value-based explanations as feature importance measures. In: Proceedings of the 37th International Conference on Machine Learning [Internet]. PMLR; 2020. p. 5491–5500. [cited 2022 May 21]. Available from: https://proceedings.mlr.press/v119/kumar20e.html [Google Scholar]
  • 81.Watson DS, Gultchin L, Taly A, et al. Local explanations via necessity and sufficiency: unifying theory and practice. In: Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence [Internet]. PMLR; 2021. p. 1382–1392. [cited 2022 May 21]. Available from: https://proceedings.mlr.press/v161/watson21a.html [Google Scholar]
  • 82.Shaikhina T, Bhatt U, Zhang R, et al. Effects of uncertainty on the quality of feature importance explanations. In: AAAI Workshop on Explainable Agency in Artificial Intelligence; Virtual; 2021. [Google Scholar]
  • 83.Hancox-Li L, Kumar IE. Epistemic values in feature importance methods: lessons from feminist epistemology. In: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency; Virtual Event Canada; 2021. p. 817–826. [Google Scholar]
  • 84.Slack D, Hilgard A, Singh S, et al. Reliable post hoc explanations: modeling uncertainty in explainability. In: Ranzato M, Beygelzimer A, Dauphin Y, et al., editors. Advances in neural information processing systems [Internet]. Curran Associates. Inc.; 2021. p. 9391–9404. Available from: https://proceedings.neurips.cc/paper/2021/file/4e246a381baf2ce038b3b0f82c7d6fb4-Paper.pdf [Google Scholar]
  • 85.Li X, Zhou Y, Dvornek NC, et al. Efficient Shapley explanation for features importance estimation under uncertainty. In: International Conference on Medical Image Computing and Computer-Assisted Intervention; Lima, Peru. Springer; 2020. p. 792–801. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 86.Barocas S, Selbst AD, Raghavan M. The hidden assumptions behind counterfactual explanations and principal reasons. In: Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency; Barcelona, Spain; 2020. p. 80–89. [Google Scholar]
  • 87.Higgins-Chen AT, Thrush KL, Wang Y, et al. A computational solution for bolstering reliability of epigenetic clocks: implications for clinical trials and longitudinal tracking. Nat Aging. 2022;2:644–661. doi: 10.1038/s43587-022-00248-2 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplemental Material

Articles from Epigenomics are provided here courtesy of Taylor & Francis

RESOURCES