Hacker News new | past | comments | ask | show | jobs | submit login
Deep biomarkers of human aging: Application of deep neural networks (impactaging.com)
71 points by jonbaer on May 22, 2016 | hide | past | favorite | 6 comments



A robust generally agreed upon and reliable biomarker of aging is a very important thing that doesn't yet exist. It would ideally require something simple like a blood sample, pin down biological age fairly accurately, and work in mice, dogs, pigs, and people at the very least. The estimate of biological age would be a measure of how damaged you are, and would have a solid statistical relationship with remaining life expectancy.

Why is this important? Because at present the only way to assess a putative rejuvenation therapy, such as the various approaches to senescent cell clearance presently under development, is to try it in mammals and wait and see. In mice that takes a lot of time and money, even when you work with one of the suppliers who can sell you pre-aged-to-the-age-you-want mice for a few hundred dollars each. Then never mind testing in longer-lived species. You'll always be able to show that your therapy does what it says on the can (removes X% of senescent cells in tissues A, B, and C) but how to prove that this does in fact extend human life in the same way it does in mice, and by how much?

Testing putative rejuvenation therapies would run a lot faster if you could just take a tissue sample a few weeks after the treatment, send it off to a clinic, and quickly see that biological age was reduced. Then far fewer of the wait and see studies would be needed, and the whole field would move a lot faster.

At present DNA methylation patterns look pretty promising as a biomarker of aging, but unfortunately there is still a way to go from the stage of pretty promising to the stage of a generally agreed biomarker of aging that is good enough to bypass the need for wait and see studies. So more signs of parallel and different approaches in the research community are always welcome.


The estimate of biological age would be a measure of how damaged you are, and would have a solid statistical relationship with remaining life expectancy.

I think a claim that aging involves only damage is reaching a bit. It seems plausible that each standard phase of human development involve adaptation as well as chromosome damage.

A standard claim is that evolve doesn't care about an organism once it is beyond the age of reproduction. Human beings live considerably longer than other primates and longer past the age of reproduction. One argument is that the extra human lifetime is an adaption to allow an aged human to aid the survival of their offspring. If this is mutation, then one might guess human genes already contain adaptions again usual genetic decay. However this may not be good news since further extending these may be harder than just eliminating the simple aging of simpler organisms.

I know that, just for example, the low metabolisms of the very elderly are seen as a protection against cancer, cancer that would otherwise be more prevalent given accumulative genetic. And this means a failure to lower metabolism over aging time might actually be associated with higher mortality (this may screw up the association of mortality and aging-process measures).


The root cause is damage. That is the mainstream view in the research community, for all that there is a lot of debate over which damage is actually fundamental and which damage is more important, or how the damage progresses in detail.

Then there is secondary and later damage caused by systematic dysregulation resulting from the root cause damage. Then there is adaptation to damage, primary and secondary and later, which is some cases is beneficial and some cases not. E.g. epigenetic changes, stem cell quiescence, cardiovascular remodeling, etc.

I'm vaguely optimistic about DNA methylation patterns as the basis for a biomarker precisely because they are not damage, but rather a reaction to that damage. In a way they are an evolved damage assessment, or at least they might be used in that way. We'll see how it pans out in practice over the next five years or so.


It also needs to be a biomarker that tracks with rejuvenation and doesn't track non-rejuvenation. Which is probably more than this result can offer.

It's like, if you have an old car, the steel body will rust in sync with the engine getting worn. But if you fix the engine, it won't be reflected in the rust, it will just break the previous association. And if you fix the rust, the engine has not been rejuvenated.


The method seems interesting and potentially useful, but I'm skeptical of the evaluation for two reasons. First, R2 is not a fair way to compare simple (linear regression, kNN) and complex (deep neural net) models, as it will always prefer a more complex model. Second, there was no indication how the error was computed. Was there a hold out set or cross-validation for accuracy? Without any more details, I have to assume they measured error on the the training data, which will also prefer the more complex model.

Also, as a public service announcement:

"The authors are affiliated with Insilico Medicine, Inc, a commercial company developing differential pathway activation scoring-based and deep learned biomarkers of multiple diseases and aging and engaging in drug discovery and drug repurposing."


I'm a fan of deep learning, but is it really the best choice in an application like this? I have always heard that shallow machine learning methods generally did better on unstructured smallish data.

They should put this data on something like Kaggle.com and let researchers and hobbyists around the world try to find the best model. At the very least publish the data somewhere.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: