How can we discover the mechanisms underlying epigenetic aging?
- Epigenetic clocks have been remarkable tools for quantifying biological aging. They are extraordinary in that the same measure can be used across diverse tissues and even diverse mammalian species. The problem is that we have no fundamental understanding of what drives DNA methylation changes with aging or how they directly connect to manifestations of aging at the tissue or organismal level. I feel that developing this understanding will require a three pronged approach—1) we need to utilize reductionistic and/or in vitro experiments to link DNA methylation changes to various aging hallmarks, gene regulatory processes, and/or other epigenetic phenomena; 2) we need to develop better single-cell data in order to link DNAm changes to cellular phenotypes and heterogeneity; 3) we need to employ computational approaches to deconstruct epigenetic clocks given that they likely capture diverse types of DNA methylation changes.
So how many samples do you need for the hypothetical study using new PC-GrimAge—looked like very few.
- Yes, according to our calculation one would only need ~100 samples to see an effect size of 1 year in PCGrimAge.
Given that different organs and tissues often have uncorrelated age acceleration, how good are the clocks that we extract from blood / saliva in predicting age acceleration in, say, the brain or bones?
- One issue is that (in humans) we don’t have a lot of methylation data measured in multiple organs from the same people. We have done some work looking at a given tissue with paired blood and it does seem that there is a lot of tissue/blood discordance. That being said, part of this disagreement may be due to which aging measures we are using. Some signals may be shared and some may diverge and it depends what you are looking at.
So do the low-noise versions of the clocks require more CpG sites to be queried?
- They do use about ~78k CpGs, but all are measured on the Illumina arrays.
How do you think about balancing age prediction and functional/phenotype prediction in aging clocks? It seems to me that there are tradeoffs between accuracy at predicting chronological age and accuracy at predicting specific aging phenotypes like cognitive decline.
- I don’t see much point in trying to optimize age prediction. Obviously you want a measure that correlates with age, but I don’t think we need to keep trying to reduce the error in age prediction. In my opinion, if you can’t predict anything above and beyond age, then why spend $$$ to get a measure (we already know age). I think the biomarker field needs a paradigm shift in which we stop focusing so much on age prediction.
Another question I have is whether the PCA technique can be used with a new dataset to create a clock that is specifically applicable to that dataset, processed in that lab, under particular conditions, with particular chips.
- For most of these clocks, the PCA and penalized regression was done in a specific dataset and then just applied to all the new validation datasets. We don’t rerun PCA in new data. Long way of saying, yes, they can be applied to new data with no issues. One can also use the method to create a brand new clock.
Will the coefficients for the PCA versions of various methylation clocks be available to researchers, or are these proprietary?
- We are working on GitHub distributions of all the PC clocks. Should be ready by the time the paper comes out.
Will the data used to derive & validate the PCA be published as well? I’ve worked with similar methods for GWAS & immune repertoire analysis, so it might be interesting to try some of them.
- We did not generate any of the data used to train the PCA clocks. They are all cohort studies that can already be applied for.
To investigate causal biology, will you be doing more non-human methylome work? Also, will you start investigating transcriptomes, proteomes, metabolomes, etc. using similar techniques but correlating between them all?
- We have a funded project to do thousands of mouse samples. Also doing a lot of intervention work in mice. We have been doing multi-omics in the brain and plan to do this in the mouse work. Once there is more human multi-omics data, that’s the direction we need to go.
How easy/feasible is it to either utilize multiple clocks from one set of samples? Or if the samples are no longer available, how easy/feasible is it to plug a set of methylation analysis into multiple clocks? Ultimately, what needs to be considered when you are designing a clinical trial so that it is easy to do? Example: If the TRIIM trial only used one clock analysis, could you do a post-hoc analysis to plug that data into more clocks to increase confidence?
- Because the human clocks are all built off arrays, one should be able to estimate all of them in the same study. One can also go back to old data and calculate new clocks that come along.