The Latest Data on Epigenetic Clocks Suggests that they are Not Yet Ready
An epigenetic clock is a weighted measure of DNA methylation at specific sites on the genome. The best such clocks correlate well with chronological age, and come with additional evidence to suggest that they also correlate well with biological age, the burden of damage that leads to dysfunction. Study populations with age-related disease, or known to have higher risks of age-related disease, also have higher ages as measured by an epigenetic clock.
Unfortunately it remains unclear as to what exactly is being measured by these epigenetic changes. They are far downstream of the damage that causes aging, and there is no clear line of cause and consequence to connect the two. That presents a challenge to those who wish to use epigenetic clocks as a way to rapidly evaluate potential rejuvenation therapies at low cost. Without knowing what the clock measures, the result is not actionable. It is quite possible that any given clock only reflects some of the root causes of aging, or some failing organ systems, and not all of them.
The results here, showing varied outcomes when epigenetic clocks are used to assess mice undergoing a variety of approaches to slow aging, suggest that epigenetic clocks are not yet ready for use in this way. Much more work remains to build clocks that can be used in confidence to quantify the performance of potential rejuvenation therapies, most of which will be bringing new mechanisms to the table, approaches that will not have been calibrated against epigenetic measures in any meaningful way.
Our understanding of age-related epigenetic changes in DNA methylation in humans has progressed rapidly with the technical advancement of genomic platforms. The correlation between chronological age and DNA methylation over the course of an entire lifespan is strong. Recent studies have taken advantage of this relationship to accurately estimate chronological age based on the methylation levels of multiple CpG dinucleotides. For example, the human multi-tissue epigenetic age estimation method combines the weighted average of DNA methylation levels of 353 CpGs into an age estimate that is referred to as DNAm age or epigenetic age.
Most importantly, we and others have shown that human epigenetic age relates to biological age, not just chronological age. This is demonstrated by the finding that the discrepancy between DNAm age and chronological age (what we term "epigenetic age acceleration") is predictive of all-cause mortality even after adjusting for a variety of known risk factors.
We combined hundreds of new DNA methylation samples collected from several mouse tissues with publicly available data from previous studies of mouse DNA methylation. We compared clocks built with different regression methods using hundreds of thousands of CpGs as input as well as a clock constructed from a limited set of mammalian-conserved CpGs. We evaluated the performance of these clocks across samples and tissues. We applied the most accurate clock to samples from previous longevity studies of mice to measure the effects of these interventions on epigenetic aging.
We demonstrate that these data enable construction of highly accurate multi-tissue age estimation methods (epigenetic clocks) for mice that apply to the entire life course (from birth to old age). We demonstrate that these clocks perform well on new tissues not included in the training of the clock by performing tissue exclusion cross-validation. This gives us confidence that these clocks will work on new samples from other tissue types as well.
Our study leads to several novel insights. First, our first prototype of an age estimator based on fewer than 1000 highly conserved CpGs demonstrates that it will be feasible to build highly accurate DNAm age estimator on the basis of highly conserved CpGs. Second, we find that epigenetic clocks that are optimal for estimating age (namely those based on elastic net regression) may be inferior to less accurate clocks (based on ridge regression) when it comes to gold standard anti-aging interventions. Only our ridge regression clock manages to corroborate most of the previously reported findings, e.g. only the ridge clock showed that dwarf strains show slower epigenetic aging relative to wild-type strains. The anti-epigenetic aging effects of calorie restriction are highly robust and could be observed with all clocks. However, none of our clocks managed to detect an anti-aging effect of rapamycin.
These results suggest that the multi-tissue ridge regression DNA methylation clock is most useful in assessing "biological age" for a variety of treatments, experimental interventions, and genetic backgrounds. However, the elastic net clocks are better for assessing chronological age. Overall, this study demonstrates that there are trade-offs when it comes to epigenetic clocks in mice. Highly accurate clocks might not be optimal for detecting the beneficial effects of anti-aging interventions.
My experience working with Ridge regressions is that the co-linearity of causal variables can be reduced or removed. I suspect there is a ton of co-linearity between the > 100,000 CpGs they looked at, some tissue specific, others strongly affecting one or several tissue types. In the end, their data set will be most useful in assessing the subset of CpGs influenced by various therapies. I wish they would have tried a genetic algorithm of combinations of and weights of CpGs against rapamycin.
Yes there is a way to go. However, the latest DNAm PhenoAge clock (513 CpGs, Levine et al 2018)represents a good progress at tracking "biological age" as constructed by regressing the phenotypic age, itself build on clinically relevant biomarkers, on DNA methylation data. Both the phenotypic age and DNAm PhenoAge are built using elastic net models though.