In this article, I’m reporting on
- a new proteomic clock from Adiv Johnson and the Stanford lab of Benoit Lehalier
- a new methylation clock developed with “deep learning” algorithms by an international group from Hong Kong
- the advanced methylation clock developed by Morgan Levine, Len Guarente, and Elysium Health
Aging clocks = algorithms that compute biological age from a set of measurable markers. Why are they interesting to us? And what makes one better than another?
The human lifespan is too long for us to do experiments with anti-aging interventions and then evaluate the results based on whether our subjects live longer. The usefulness of an aging clock is that it allows us to quickly evaluate the effects on aging of an intervention, so we can learn from the experiment and move on to try a variant, or something different.
Many researchers are skeptical about using clock algorithms to evaluate anti-aging interventions. I think they are right to be asking deep questions; I also think that in the end the epigenetic clocks in particular will be vindicated for this application.
It may seem obvious that we want the clock to tell us something about biological aging at the root level. We are entranced by the sophisticated statistical techniques that bioinformaticists use to derive a clock based on hundreds of different omic factors. But all that has to start with a judgment about what’s worth looking at.
Ponder this: The biostatisticians who create these clocks are optimizing them to predict chronological age with higher and higher correlation coefficient r. But if they achieve a perfect score of r=1.00, the clock becomes useless. It cannot be used to tell a 60-year-old with the metabolism of a 70-year-old from another 60-year-old with the metabolism of a 50-year-old, because both will register 60 years on this “perfect” clock.
It’s time to back up and ask what we think aging is and where it comes from, then optimize a clock based on the answer. As different people have different answers, we will have different clocks. And we can’t objectively distinguish which is better. It depends on whose theory we believe.
Straw man: AI trained to impute age from facial photos now has an accuracy of about 3½ years, in the same ballpark with methylation clocks. If we used these algorithms to evaluate anti-aging interventions, we would conclude that the best treatments we have are facelifts and hair dye.
Brass tacks: People with different positions about the root cause of aging all agree that (a) aging manifests as damage, and (b) methylation and demethylation of DNA take place under the body’s tight and explicit site-by-site regulation.
But what is the relationship between the methylation and the damage? There are three possible answers.
- (from the “programmed” school) Aging is programmed via epigenetics. The body downregulates repair mechanisms as we get older, while upregulating apoptosis and inflammation to such an extent that they are causes of significant damage.
- (from the “damage” school) The body accumulates damage as we get older. The body tries to rescue itself from the damage by upregulating repair and renewal pathways in response to the damage.
- (also from the “damage” school) Part of the damage the body suffers is dysregulation of methylation. Methylation changes with age are stochastic. Methylation becomes more random with age.
My belief is that (1), (2), and (3) are all occurring, but that (1) predominates over (2). The “damage” school of aging would contend that (1) is excluded, and there are only (2) and (3).
How can these three types of changes contribute to a clock?
(3) makes a crummy clock, because, by definition, it’s full of noise and varies widely from person to person and from cell to cell. There is no dispute that a substantial portion (~50%) of age-related changes in DNA methylation are stochastic. But these changes are not useful and, in fact, most of the algorithms used to construct methylation clocks tend to exclude type (3) changes. I won’t say anything more about stochastic changes in methylation, but I’ll acknowledge that there is more to be said and refer you to this article if you’re interested in methylation entropy.
If you are from the “damage” school, you don’t believe in (1), so this leaves only type (2). If changes in methylation are the body trying to rescue itself, then any intervention that makes the body’s methylation “younger” is actually dialing down protection repair. You expect that reducing methylation age will actually hasten aging and shorten life expectancy. You have every reason to distrust a clinical trial or lab experiment that uses methylation age as criterion for success.
|White cell count is used as a reliable indication of cancer. As cancer progresses, white cell count increases. The higher a person’s white cell count, the closer he is to death. So let’s build a “cancer clock” based on white blood count, and let’s use it to evaluate anti-cancer interventions. The best intervention is a chemical agent that kills the most white blood cells. It reliably sets back the “cancer clock” to zero and beyond. But we’re puzzled when we find that people who get this intervention die rapidly, even though the cancer clock predicted that they were completely cured. The problem is that white blood cells are a response to cancer, not its cause.|
If you are from the “programmed” school, you think that (1) predominates, and that a clock can be designed to prefer type (1) changes to (2) and (3). Then methylation clocks measure something akin to the source of aging, and we can expect that if an intervention reduces methylation age, it is increasing life expectancy.
The fact that methylation clocks trained on chronological age alone (with no input concerning mortality or disease state) turn out to be better predictors of life expectancy than age alone is a powerful validation of methylation technology. But only if you believe (for other reasons) that methylation is an upstream cause of aging. You could expect this from either type (1) or type (2) methylation changes.
I believe that aging is an epigenetic life program, and that methylation is one of several epigenetic mechanisms by which it is implemented. That’s why I have faith in methylation clock technology.
Conversely, people who believe that the root cause of aging is accumulated damage are right to discount evidence from epigenetic clocks as it pertains to the efficacy of particular treatments. As in the cancer example above, treatments that create a younger methylation age can actually be damaging.
The basis for my belief that aging is an epigenetic program is the subject of my two books, and was summarized several years ago in this blog. I first wrote about methylation as a cause of aging in this space in 2013. For here and now, I’ll just add that we have direct evidence for changes of type (1). Inflammatory cytokines are up-regulated with age. Apoptosis is upregulated with age. Antioxidants are downregulated with age. DNA repair enzymes and autophagy enzymes and protein-folding chaperones are all down-regulated with age. All these are changes in gene expression, presumably under epigenetic control.
Which is more basic, the proteome or the methylome?
For reasons I have elaborated often in the past, I adopt a perspective on aging as an epigenetic program. I think of methylation clocks as close to the source, because methylation is a dispersed epigenetic signal. But the proteome is, by definition, the collection of all signals transmitted in blood plasma, including all age signals and transcription factors that help to program epigenetics cell-by-cell. The proteome is generated by transcription of the DNA body-wide, which transcription is controlled by methylation among other epigenetic mechanisms. So one might argue from this that the methylome is further upstream than the proteome. On the other hand, methylation is just one among many epigenetic mechanisms, and the proteome is the net result of all of them. On this basis, I would lean toward a proteomic clock as being a more reliable surrogate for age in clinical experiments, even better than methylation clocks. It is a historic fact, however, that methylation clocks have a 6-year headstart. Methylation testing is entering the mainstream, with a dozen labs offering individual readings of methylation age, priced to attract end-users.
Let’s see if proteomic clocks can catch up. The new technology is based on SOMAscan assays, and so far is marketed to research labs, not individuals or doctors, and it is priced accordingly. The only company providing lab services is SOMAlogic.com of Boulder, CO. “SOMAscan is an aptamer-based proteomics assay capable of measuring 1,305 human protein analytes in serum, plasma, and other biological matrices with high sensitivity and specificity.” [ref] As I understand it, they have a microscope slide with 1305 tiny dots, each containing a different aptamer attached to a fluorescent dye. An aptamer is like an engineered antibody, optimized by humans to mate to a particular protein. Thus 1305 different proteins can be measured by applying a sample (in our case, blood plasma) to the slide, chemically processing the slide to remove aptamers that have not found their targets, then photographing the slide and analyzing the readout from the fluorescent dye.
Aptamers are synthetic molecules that can be raised against any kind of target, including toxic or non immunogenic ones. They bind their target with affinity similar or higher than antibodies. They are 10 fold smaller than antibodies and can be chemically-modified at will in a defined and precise way. [NOVAPTech company website]
Curiously, aptamers are not usually proteins but oligonucleotides, cousins of RNA, simply because the chemical engineers who design and optimize these structures have had good success with the RNA backbone. The SOMA in SOMAlogic stands for “Slow Off-rate Modified Aptamers”, meaning that the aptamers have been modified to make them stick tight to their target and resist dissociating.
An internal proteome-methylome clock?
It’s possible that there is a central clock that tells the body “act your age”. I have cited evidence that there is such a clock in the hypothalamus, and that it signals the whole body via secretions [2015, 2017].
Another possibility is a dispersed clock. The body’s cells manufacture proteins based on their epigenetic state, the proteins are dispersed in the blood, some of these are received by other cells and affect the epigenetic state of those cells. This is a feedback loop with a whole-body reach, and it is a good candidate for a clock mechanism in its own right.
|I’m interested in the logic and the mathematics of such a clock in the abstract. Any feedback loop can be a time-keeping mechanism. Such a mechanism is
_____Epigenetics ⇒ Protein secretion ⇒ Transcription factors ⇒ Epigenetics
This is difficult to document experimentally, but it is an attractive hypothesis because it would explain how the body’s age can be coordinated system-wide without a single central authority, which would be subject to evolutionary hijacking, and might be too easily affected by individual metabolism, environment, etc. But the body’s aging clock must be both robust and homeostatic. If it is thrown off by small events, it must return to the appropriate age. So my question—maybe there are readers who would like to explore this with me—is whether it is logically possible to have a timekeeping mechanism that is both homeostatic and progressive, without an external reference by which it can be reset.
Last year, Lehalier and a Stanford-based research group jumpstarted the push toward a methylomic aging clock with this publication [my write-up here]. The same group has a follow-up, published a few weeks ago. The new work steps beyond biologically agnostic statistics to incorporate information about known functions of the proteins that they identified last year. The importance of this is twofold: It suggests targets for anti-aging interventions. And it supports the creation of a clock composed of upstream signals that have been verified to have an effect on aging. I argued in the long Prelude above that this is exactly what we want to know in order to have confidence in an algorithmic clock as surrogate to evaluate anti-aging interventions.
They work with a database I had not known about before: the Human Ageing Genomic Resources Database. HAGR indexes genes related to aging and summarizes studies that document their functions. Some highlights of the proteins they identified:
- Inflammatory pathways are right up there in importance. No surprise here. But if you can use inflammatory epigenetic changes to make an aging clock, you have a solid beginning.
- Sex hormones that change with age turn out to be even more prominent in their list. The first several involve FSH and LH. These are hormones connected with women’s ovarian cycles; but after menopause, when they are not needed, their prominence shoots up, and not just once-a-month, but always on. Men, too, show increases in LH and FSH with age, though they are more subtle. I first became aware of LH and FSH as bad actors from the writings of Jeff Bowles more than 20 years ago.
- “GDF15 It is a protein belonging to the transforming growth factor beta superfamily. Under normal conditions, GDF-15 is expressed in low concentrations in most organs and upregulated because of injury of organs such as such as liver, kidney, heart and lung.” [Wikipedia] “GDF15 deserves a story of its own. The authors identify it as the single most useful protein for their clock, increasing monotonically across the age span. It is described sketchily in Wikipedia as having a role in both inflammation and apoptosis, and it has been identified as a powerful indicator of heart disease. My guess is that it is mostly Type 1, but that it also plays a role in repair. GDF15 is too central a player to be purely an agent of self-destruction.” [from my blog last year]
- Insulin is a known modulator of aging (through caloric restriction and diabetes).
- Superoxide Dismutase (SOD2) is a ubiquitous antioxidant that decreases with age, leaving the body open to ROS damage.
- Motilin is a digestive hormone. Go figure. Until we understand more, my recommendation would be to leave this one out of the aging clock algorithm.
- Sclerostin is a hormone for bone growth. It may be related to osteoporosis, and well worth inclusion.
- RET and PTN are called “proto-oncogenes” and are important for development, but associated with cancer later in life.
Which proteins are most relevant?
The Horvath clocks have been created using “supervised” optimization, which involves human intelligence that oversees the application of sophisticated algorithms. But what happens if you automate the “supervised” part? On the one hand, you must expect mistakes and missed opportunities that you wouldn’t have with human supervision. On the other hand, once you have a machine learning algorithm, you can apply it over and over to different subsets of the data, produce hundreds of different clocks, and choose those that perform best. That’s what Johnson and co-authors have done in the current paper. They describe creating 1565 different clocks based on different subsets of a universe of 529 proteins. In my opinion, their most important work combines biochemical knowledge with statistical algorithms. The work using statistical algorithms alone are much less interesting, for reasons detailed in the Prelude above.
This new offering from Lehalier and Johnson is a great step forward in that
- proteins in the blood are a broader picture of epigenetics than methylation alone
- specific proteins are linked to specific interventions that are reliably connected to aging in the right direction. Crucially, the clock is designed to have type (1) epigenetic changes (from the Prelude above) and to exclude type (2)
- to calibrate the clock not with calendar age but with future mortality. This would require historic blood samples, and it is the basis of the Levine/Horvath PhenoAge clock.
- to optimize the clock separately for different age ranges or, equivalently, to use non-linear fitting techniques in constructing the clock algorithm
- to commercialize the Aptomer technology, so that it is available more widely and more cheaply
Elysium is a New York company advised by Leonard Guarente of MIT and Morgan Levine (formerly Horvath’s student, now at Yale). They have an advanced methylation clock available to the public, which they claim is more accurate than any so far. Other clocks are based on a few hundred CpG sites that change most reliably with age, but the Index clock uses 150,000 separate sites (!) which, they claim, offers more stability. The Horvath clocks can be overwhelmed by a single CpG site that is measured badly. (I have personal experienc with this.) Elysium claims that variations from one day to the next or one lab slide to the next tend to average out over such a large number of contributions. On the other hand, as a statistician, I have to wonder about deriving 150,000 coefficients from a much smaller number of inividuals. The problem is called overfitting, and the risk is that the function doesn’t work well outide the limited data set from which it was derived.
In connection with the DataBETA project, I have been talking to Tina Hu-Seliger, who is part of the Elysium team that developed Index. I am impressed that they have done some homework that other labs have not done. They compare the same subject in different slides. They store samples and freeze them and compare results to fresh samples. They compare different clocks using saliva and blood.
I wish I could say more but Elysium Index is proprietary. There is a lot I have not been told, and there is more that I know that I have been asked not to reveal. I don’t like this. I wish that all aging research could be open sourced so that researchers could learn from one another’s work.
Two other related papers
DeepMAge is a new methylation clock, published just this month, based on more sophisticated AI algorithms instead of the standard 20th-century statistics used by Horvath and others thus far. Galkin and his (mostly Hong Kong, mostly InSilico) team are able to get impressive accuracy in tracking chronological age. This technology has forensic applications, in which evidence of someone’s calendar age is relevant, independent of senescence. And the technology may someday be the basis for more accurate predictions of individual life expectancy. But, as I have argued above, a good clock for evaluating anti-aging measures must look at more than statistics. Correlation is not the same as causation, and only detailed reference to the biochemistry can give confidence that we have found causation.
Biohorology is a review paper from some of this same InSilico team together with some prominent academics, describing the latest crop of aging clocks. The ms is long and detailed, yet it never addresses the core issue that I raise in the Prelude above, about the need to distinguish upstream causes of aging from downstream responses to damage.
The beginning of the ms contains a gratuitous and outdated dismissal of programmed aging theories.
“Firstly, programmed aging contains an implicit contradiction with observations, since it requires group selection for elderly elimination to be stronger than individual selection for increased lifespan.”
Personally, I bristle at reading statements like this. which ignore an important message of my own work and, more broadly, ignore the broadened understanding of evolution that has emerged over the last four decades.
“Secondly, in order for the mechanism to come into place, natural populations should contain a significant fraction of old individuals, which is not observed either (Williams, 1957).”
This statement was the basis not just of Williams’s 1957 theory, but more explicitly of the Medawar theory 5 years earlier. Neither of these eminent scientists could have known that their conjecture about the absence of senescence in the wild would be thoroughly disproven by field studies in the 1990s, The definitive recent work on this subject is [Jones, 2014].
For the purpose of evaluating anti-aging treatments, the ideal biological clock should be created with these two techniques:
- It should be trained on historic samples where mortality data is available, rather than current samples where all we know is chronological age, and
- Components should be chosen “by hand” to assure all are upstream causes of aging rather than downstream responses to damage. (Type 1 from analysis above.)