Methylation update, Part II
Imagine Horvath’s thought process last year, when the PhenoAge clock (described last week) was derived. In order to evaluate anti-aging interventions in humans, the most useful measure would be a clock that estimates not how many years since your birth but how many years until your death. The 2013 methylation clock and the (non-methylation) blood tests combined to create PhenoAge both did a good job, and there was little overlap between the two. So combining an epigenetic/methylation measure with non-methylation blood tests might be the basis for an even more accurate estimate of time-to-death. There are also life-style factors that could be factored in, e.g., smoking, diet, exercise, socio-economic status.
Last spring, Horvath set his insightful project scientist, Ake Lu, to work on their “GrimAge” clock (named after the grim reaper). But a funny thing happened on the way to the spreadsheet. They started with a large training set of 2400 blood samples from the Framingham Heart Study, which has been collecting data since 1948. They supplemented the methylation data with blood markers and the known smoking history of each patient to create a composite index. The next step was standard statistical procedure: quantifying the overlap between the methylation and non-methylation data to eliminate redundancy. For example, they asked: to what extent is smoking history already reflected in methylation status? The surprising result was that the methylome already knew all about the smoking history and the body’s response to it. In fact, the methylation sites associated with smoking history predicted how long the person would live more accurately than the smoking history itself.
Remember from last week that the PhenoAge methylation clock was derived from the PhenoAge blood markers, and that the methylation version did not do as good a job at predicting mortality as the blood markers from which it was derived. This is the expected situation.
But this time, Horvath and Lu were confronted with a case where the information they had hoped to use to supplement methylation data was actually reflected in (different) methylation data, and the reflection worked better than the original. The methylation changes–presumably a response to smoking–told more about each person’s health risk than did the smoking history itself. Even stranger, the methylation marks most closely associated with smoking were found to be a powerful indication of future health even when the sample was confined to non-smokers.
If they continued undeterred on their original plan to add smoking status as a health indicator alongside methylation status, then the coefficient for smoking would have to be positive; yes, the math was telling them that, after allowing for all the information in the methylation profile, the extra information that a person had been a heavy smoker would actually lengthen the estimate of life expectancy, after the methylation response to smoking had been taken fully into account.
What could this possibly mean? Lu and Horvath don’t speculate on this point, but here are the three possibilities I can think of:
- Smokers are not reporting their history accurately, perhaps from shame or from censored memory. The methylation response is actually a better indication of the number of pack-years smoked than the person’s memory of the number of pack-years.
- The lung damage by smoking is highly individual. Each person’s response to smoking depends both on the number of cigarettes smoked and also his susceptibility to damage, and these two factors are reflected in the methylation pattern, which is a response to smoking.
- Most radical of all is the possibility that smoking kills not directly by damaging the lungs and arteries, but indirectly by inducing the body to alter gene expression toward an older, less healthy state. Radical, yes, but the only one of these three ideas that might explain why the methylation patterns predict mortality in non-smokers.
Rather than continue with this perverse conclusion, Lu and Horvath pursued their analysis with redoubled respect for the power of methylation indicators to predict age and age-related health. They looked for other markers–blood levels of certain proteins that might supplement methylation data in their Grim Age clock. And they found the same phenomenon as with the smoking. Yes, the blood markers held information about the individual’s future health prospects, but each marker also had its image in the DNA methylation pattern, and in several other cases (e.g. PAI-1 and TIMP-1) the methylation based surrogate marker was a better predictor of lifespan than was the original plasma protein level from which it was derived.
Some of these proteins will sound familiar to aging researchers: GDF15=Growth differentiation factor 15 (which should not be confused with GDF11). CRP=C-Reactive Protein, is a well-recognized marker of inflammation, which contributes to all diseases of old age. Others are more obscure. Cystatin-C is a blood marker of kidney function that more recently has been found to be a robust predictor of cardiovascular outcomes. TIMP1 is a protein that displays an impressively tight correlation with age, but I couldn’t begin to describe its biochemical function.
The article calls attention to the gene PAI-1, which I had never heard of. Plasma Activator-Inhibitor 1, aka, SERPIN-E1, regulates blood clotting, which is an important contributor to heart attacks and stroke. Later in life, de-methylation of suppressor regions in a chromosome causes more PAI-1 to appear in the blood, leading to increased heart risk. For no apparent reason, PAI-1 turns out to be a powerful predictor of heart disease, diabetes, fatty liver, and of age-related disease in general.
I would have liked to see correlation coefficients for all these measures because p values get better with more data, even if the correlation is weak. r tells you how much scatter you can expect if you try to extract information from the methylation profile of an individual or group of individuals in the future, but p only reassures you that yes, the correlation is not the result of chance. Horvath responded to me that there are technical reasons that r values cannot be inferred directly using the kinds of data on which his calculations were based.
Direct vs Indirect
Here’s another paradox. The DNAm GrimAge clock was developed in two stages, a correlation of a correlation. How does it compare to a direct, single stage computation of the methylation pattern that best predicts mortality (in technical language: a linear regression of time to death on the methylation profile)? In the Supplemental Materials published online with GrimAge, Horvath and Lu compare their GrimAge clock to Zhang’s clock (see last week) and to their own single-stage computation, developed for this purpose. Curiously, the indirect computation yields the better result. Why? In an email message, Horvath said he is just as surprised and puzzled by the result as I am.
An implication for Anti-Aging Lifestyle
Aside from the corroboration that we shouldn’t smoke cigarettes (duh), there is just one other direct implication for lifestyle in the GrimAge paper. They report longer life expectancies for people taking omega 3 supplements. The effect was on the edge of statistical significance, and more pronounced in men than in women. But it corroborates results from human epidemiology. A word to the wise.
Why the methylation clock is able to detect omega 3 supplements is again puzzling. We imagine that omega 3 in the diet acts directly on the lipids in the bloodstream, and that is where the health benefits come from. But it seems that dietary omega 3 affects the methylome as well. If this were just a response to the blood lipids, we would not expect it to correlate so well with the aging clock. Once again, the methylation clock is proving more robust than even its proponents would have guessed.
Methylation clocks to evaluate life extension technology
I have been enthusiastic about the potential of methylation clocks to screen life extension interventions and tell us what works. In fact, I’m organizing a trial in humans to test many common interventions and their interactions. If we think of the methylation clock as a faster, cheaper replacement for lifespan statistics, then the DNAm GrimAge clock is the latest and greatest tool we have. It is thus important to ask, what is the evidence for a close correspondence between interventions that slow the methylation clock and interventions that lengthen life expectancy? In short, there is evidence of a close but not perfect correspondence. I reviewed the evidence last year,
Eating red meat shortens life expectancy, and indeed it increases GrimAge. Conversely, vegetables, nuts, and fruits in the diet increase life expectancy and they lower GrimAge. HDL levels in the blood are good for longevity and lower GrimAge. Markers of inflammation are associated with faster aging, and also with higher GrimAge. Blood sugar control is important for longevity, and it appears to be reflected in GrimAge. Perhaps less expected, higher levels of education and income are associated with longer life expectancy, and both seem to be robustly mirrored in methylation, as measured by GrimAge. Age acceleration from smoking is well-reflected in GrimAge. Early menopause forbodes an early death, and this, too, has fingerprints in GrimAge.
On the other hand, we think rapamycin is the best candidate yet for an anti-aging drug, and no significant effect of rapamycin on methylation age has yet been detected. Obesity is associated with life shortening, but only weakly accelerates GrimAge. Aspirin, metformin, and vitamin D are supplements that are thought to have a small but significant benefit for lifespan. Do the methylation clocks pick up these effects? I have not seen data that they do. The fact that telomerase expression seems to accelerate methylation clocks gives pause.
And this study provides grounds for caution. Blood stem cells from the bone marrow were transplanted for medical reasons, and years later, the blood cells derived from the donor stem cells were collected and analyzed for methylation age. The result was that the blood cells remembered the age of the donor. They were not re-programmed by the new environment to match the age of the recipient’s body. While this result can’t detract from the accuracy of aging clocks based on methylation, it raises a theoretical and a practical issue. The result weighs against a theory (which has been a favorite of mine) that aging is programmed centrally, and that information about the body’s age is transmitted throughout the body by signals in the blood plasma. And it also calls into question the assumption (at the root of my Data-BETA study) that methylation clocks based on the blood will respond with the body if an anti-aging intervention is effective.
Other applications—other clocks
GrimAge takes the prize as the best candidate to replace the lifespan study, which is our current gold standard for evaluating anti-aging interventions.But there remain other uses for methylation clocks, and there is every reason to develop other clocks which predict other aspects of aging:
- Brain aging–perhaps a composite of reaction time and ability to form new memories
- Fast twitch muscles for sprinting
- Mitochondrial efficiency and aerobic capacity
- Cardiovascular age, from loss of elasticity in artery walls and stiffening of the heart muscle with glycation
- Aging of the immune system
The Bottom Line
Horvath and Lu have given us the most accurate epigenetic predictor yet of future mortality and morbidity, and, surprisingly, it is based in methylation alone, and not the other blood markers and lifestyle factors that they had originally thought would supplement methylation. Horvath’s finding that secondary methylation indicators are more accurate than the underlying primary indicator from which they were derived is provocative, and calls out for a new understanding. It suggests that methylation clocks might be even more robust than we thought. On the other hand, the recent finding that blood stem cells transplanted from one body into another retain a memory of the donor’s age suggests just the opposite.