I recently read through a survey of AI and aging/longevity research, and there were some exciting tie-ins to my research to takeaway, plus some generally cool stuff going on. The survey is titled Artificial intelligence for aging and longevity research: Recent advances and perspectives. The paper is geared towards a medical crowd rather than an ML one, and the main idea is to present the state of machine learning within the aging and longevity research space. The review briefly covers where has ML succeeded, what problems are most pressing, where ML might help the most, which I will recap quickly in this post.

The end-goal of machine learning in aging and longevity research is to enable researchers to disentangle the overwhelming amount of available data, develop personalized interventions, and more accurately quantify the aging process.

What can ML researchers take away from this paper?

1) Medical ML needs more interpretable methods

A problem like aging is just so complex. We really want to understand more about the process of aging itself, and for that, we need our various ML models to be able to tell us what they see, and what’s going on biologically as people age.

We don’t just want to ask the model: “How old is Andrew?”, we want to ask it: “Which factors in Andrew’s profile contribute to your estimation of his age?” More interpretable models will help researchers make sense of massive amounts of data, and provide more concrete benefits to medical research, including guidance on where to allocate further research effort and funding, candidates compounds within drug discovery, biomarkers for cancers and other diseases, and more.

2) ML can help us to better predict age

While we don’t just want to ask models: “How old is Andrew?”, we definitely still want to ask that question! Age is very difficult to predict, and there is a distinction between how old somebody actually is (in years) and how old their body seems to be. One crucial hurdle in aging research is just understanding how old somebody is! If we can get meaningful, comparable estimates of age, then we can better understand what makes two people different and how aging affects biological processes.

3) Medical ML can speed up long-term studies

A fundamental roadblock in aging research is simply that it takes so much time. We can’t easily make a change in somebody’s life and then measure how much that change has affected their aging process, as such changes are not visible for years or decades. Modern generative machine learning could help to mitigate this issue. If we can learn to plausibly and realistically forecast an individual’s age, we could generate long-term aging data without needing to wait decades and meticulously gather data.

4) References to datasets

The paper includes links and citations for several medical datasets which may be of interest to researchers looking to get into medical ML! Owing to privacy concerns, medical data is often difficult to come by in this space. I’ve included links to all of the datasets at the end of this post.

Being able to forecast or forward-and-backward simulate an individual’s age would help to predict the effects of different drugs, lifestyle changes, or vitals on a person’s aging process.

Closing Thoughts & Discussion

While I’m very interested in medical machine learning, one of my primary focus areas is interpretability. It’s great to see that interpretability will have a tangible benefit to society, and that it isn’t just some weird abstract problem that researchers chase for no real audience.

I’m also excited by the prospects of medical research being accelerated by ML, and ways that medical ML could help benefit the rest of the ML community. For example, the survey raises interesting problems such as “how do we accurately represent molecules for deep networks, while still maximizing the signal-to-noise ratio of our network inputs?” If we can solve this problem in medicine, surely we will have an interesting solution for graph neural networks throughout ML.

I’d love to see reinforcement learning and interpretability researchers pick up aging as a demonstration domain. It isn’t too difficult to imagine a pipeline where an agent begins with a person’s profile and then a “game engine” forward-simulates that person’s aging process, and the agent tries to minimize the person’s biological age throughout the episode. Which drugs, introduced when, have the greatest impact? Where should vitals be? At what time? Obviously, some pieces of the pipeline are missing, but it doesn’t sound too outlandish!

Citations for the survey paper and datasets referenced (with descriptions from the paper):

Zhavoronkov, Alex, et al. "Artificial intelligence for aging and longevity research: Recent advances and perspectives." Aging research reviews 49 (2019): 49-66.
Tacutu, Robi, et al. "Human Ageing Genomic Resources: new and updated databases." Nucleic acids research 46.D1 (2018): D1083-D1090.
Moskalev, Alexey, et al. "Geroprotectors. org: a new, structured and curated database of current therapeutic interventions in aging and age-related disease." Aging (Albany NY) 7.9 (2015): 616.
Moskalev, Alexey, et al. "Aging Chart: a community resource for rapid exploratory pathway analysis of age-related processes." Nucleic acids research 44.D1 (2016): D894-D899.
Budovsky, Arie, et al. "LongevityMap: a database of human genetic variants associated with longevity." Trends in Genetics 29.10 (2013): 559-560.
Wilmoth, John R., et al. "Methods protocol for the human mortality database." University of California, Berkeley, and Max Planck Institute for Demographic Research, Rostock. URL: http://mortality. org [version 31/05/2007] 9 (2007): 10-11.
CellAge— “A manually curated database of genes associated with cell senescence”: https://genomics.senescence.info/cells/
LongevityMap— “a repository of genetic association studies of longevity…”: https://genomics.senescence.info/longevity/
GenAge— “a benchmark database of age-related genes.”: https://genomics.senescence.info/genes/
Human Ageing Genomic Resources— source for the above: https://genomics.senescence.info/index.php
Geroprotectors— “a curated database of geroprotectors.”: https://geroprotectors.org/
AgingChart— “a list of pathways implicated in aging and longevity”: http://agingportfolio.org/
The Human Mortality Database— “information to analyze demographic trends including mortality and fertility rates.”: https://www.mortality.org/