Computational Aging

One of the biggest challenges with slowing aging and improving health is that we don't closely measure how we age. Most people do annual blood tests, look for changes in how they look in the mirror, and wait until they get a specific disease. This is like taking basic measurements once a year or even less. It means our understanding of health and aging is very limited. We can't quickly try new things and see if they work. New therapies are beholden to the pace of slow, expensive clinical trials that take decades to bring new therapies to market. At this pace, we will have to get lucky to cure aging. In short, today's healthcare system leads to slow, incremental, analog improvements. 

Instead, we should track our complete health real-time. We should have real-time, continuous, zero-marginal-cost data on the health of every tissue and system, our rate of molecular damage throughout our bodies, and frequent predictions on our mounting health risks. It would provide tremendous understanding of aging and health. Then, we would apply modern ML tools to this data across lots of people with lots of health outcomes over time. These ML tools would be able to model our health, predict our health risks, and suggest interventions to improve our health. It's hard to overstate what this will bring.

Imagine a Chat-GPT for our own personal health. Let's call it Doc-GPT. Doc-GPT would be trained on data from hundreds of millions of people over many years. It will comb your data looking for risks and making suggestions. You will get reminders and suggestions frequently based on your risks and behavior. It will become your primary preventative physician.

I believe we are trending to this world. We will have automatic, continuous measurement of aging and health. Two big trends are making this possible: 1) the explosion of health data tracked on our phones and wearables, and 2) the exponential decrease in the cost of 'omics measurements. My iPhone and Apple Watch already have the data to make much better sense of my health than any doctor. They know my heart rate, gait, voice, daily activities, and much more. Our heart rate variability, voice, and gait each provide windows into our health and molecular damage. They have my photo library from the last 10 years, and faces provide a good window into our overall health [1]. They know how social and active we are, and how much we exercise. They know when people have heart attacks. They know when we go to the hospital. They can sense dementia. They probably know when we die. Apple has over 1 billion iPhone users and 100 million Watch users. They can see what increases or slows our aging. They can see what helps and hurts our health.

Our wearables and phones don't yet directly know the molecular damage in our cells and DNA. This is fundamental data to aging and health. Fortunately, 'omics measurement tools are riding an exponential curve trending towards essentially zero cost. ~20 years ago it cost billions of dollars to map the human genome. Today, we can do it for around $100. This is trending towards zero. As it does, we will know the status our DNA, epigenetics, RNA, and proteins much more frequently. It's hard to imagine having this data in real-time, but we are trending in that direction.

This health data from phones, wearables, and omics is absolutely massive. It's on the scale to build the Doc-GPT of the future. Apple, for example, may already be able to build a Doc-GPT with the data they have that is better than any doctor today. These data may replace much of what clinical trials do. These tools are coming, and they can't come soon enough. The race is on.

References:

[1] Xia X, Chen X, Wu G, Li F, Wang Y, Chen Y, Chen M, Wang X, Chen W, Xian B, Chen W, Cao Y, Xu C, Gong W, Chen G, Cai D, Wei W, Yan Y, Liu K, Qiao N, Zhao X, Jia J, Wang W, Kennedy BK, Zhang K, Cannistraci CV, Zhou Y, Han JJ. Three-dimensional facial-image analysis to predict heterogeneity of the human ageing rate and the impact of lifestyle. Nat Metab. 2020 Sep;2(9):946-957. doi: 10.1038/s42255-020-00270-x. Epub 2020 Sep 7. PMID: 32895578.   

In 2022, I started earnestly working in biomedicine after over a decade in tech and startups. As the year nears a close, I wanted to reflect on my work and the general outlook for aging science.

Is aging science the right focus? My big goal is that humanity achieves significant lifespan and healthspan extension by 2063. I decided to direct my professional energy to this, while maintaining a high quality of life otherwise. To be clear, It does not mean sacrificing today's life for some future life. Overall, I still think this is a great goal. By 2050, 1.3 billion people will be over 65 years old [1] and likely starting to suffer from age-related decline and disease. It would be great personally, for my family and friends, and for billions of people around the world to have more healthy years.

Is the goal realistic? At a high-level, yes. We know that biology figured out how to reproduce young cells and organisms from old cells via reproduction. There are "negligibly senescent" creatures like the hydra, ocean quahog, Galapagos tortoise, and naked mole rat. I've been inspired again and again by new discoveries this year. Increasingly, it seems that cellular reprogramming can reduce the functional age of cells and tissues without causing cancer. Cellular reprogramming on its own could be one of the defining technologies for the 21st century. The advances in machine learning have been stunning, from AlphaFold to Chat-GPT. These will increasingly be applied to biomedical discovery. Of critical importance, there is no shortage of funding for this work in the foreseeable future. The aging population, which happens to hold most of societal wealth, will continue to invest more and more money into biomedicine, which will attract new talent and ideas, and encourage faster progress. There will be millions and millions of brilliant people around the world working on aging science.

What should the next decade look like for aging science? We still haven't proven the fundamental science. There are increasingly powerful theories of aging (e.g. Sinclair's "information theory"), but we still have open questions of what drives aging. Cellular reprogramming is incredibly exciting, but it still needs to be proven to safely extend long-term lifespan in other organisms, let alone humans. Priority #1, in my view, is to progress fundamental aging science as quickly as possible, including cellular reprogramming. Priority #2 should be bringing discoveries from the last few decades to drugs that can moderately extend healthspan and lifespan. Senolytics are compelling for moderate life extension, but we don't yet have proven drugs. Young blood plasma is surprisingly powerful, yet we don't have safe therapies. Rapamycin, metformin, and new diet and exercise interventions are all currently in clinical trials. We need interventions like these to start hitting the market in the next decade, which will extend healthspan and build momentum for the field. Priority #3 should be exploring new, risky ideas for aging science, beyond the current mainstream. How do other organisms achieve negligible senescence? How can advances in machine learning and large-language models reveal insights into aging?

I'm thrilled to be working in this field in an actual paying job. The learning curve is steep. I hope to contribute by my own efforts in science. I also hope to inspire others to switch their careers into aging science.

References:

[1] United Nations Department of Economic Social Affairs. World population prospects 2019: highlights. New York: United Nations Department of Economic Social Affairs; 2019.

After over a decade in startups, I recently joined UCSF’s Abbasi Lab. I began in February 2022 as a volunteer, and then I went full-time in September 2022 on a one-year specialist appointment. It has been a wonderful experience so far, and I wanted to share why.

My big goal and hope is that humanity significantly slows or even reverses aging in the next ~40 years. In exploring how to have maximum impact, I am pursuing the computational / data science path vs. the wet lab / biological sciences path. It's a better fit personally. Biological data collection and machine learning are each improving rapidly, likely exponentially. I learned from the startup world that it’s good to be a part of things that are improving exponentially. The marriage of the two will lead to amazing discoveries in the coming years. In my work, I hope to build better ways to measure human aging. With better measures, we can more quickly figure out how to slow aging. Abbasi Lab was a natural fit, given it is a computational and machine learning (ML) lab working on biological data. I also like that Abbasi Lab has a focus on neuroscience––if we can’t slow brain aging, then it probably won’t be worth having longer lifespans.

The work is cutting edge and highly relevant to aging science. My focus is currently applying machine learning to spatial transcriptomics (ST) data. ST allows measurement of the expression (transcripts) of 1000s of genes in their native spatial location at single-cell resolution. In 2021, ST was named “Method of the Year” by Nature Methods [1] because of the potential for spatially-resolved gene expression data to unlock new secrets about biology, development, disease, and aging. The sheer size and rapid growth of ST datasets requires ML to make sense of it. Like ST, ML has seen incredible method development in recent years. ML is beginning to solve previously-flummoxing biological problems, such as Google’s AlphaFold for protein folding [2]. However, ML can be especially challenging to apply to biology, as the results from ML models need to be accurate, repeatable, and interpretable to biological reality [3], [4]. Thus, we are developing accurate, repeatable, interpretable ML tools and frameworks to help biologists explore and analyze ST data.

Other Abbasi Lab projects are also breaking new ground at the intersection of biological data and ML. One of our team members is creating a deep learning system to take medium resolution MRI images and automatically turn them into high resolution MRI images, which can more accurately measure the progression of brain disease and aging. Another is building an ML system to diagnose severity of Parkinson’s Disease using only brief videos of patients walking. One that I am particularly excited about is an effort to measure brain disease and related functional phenotypes based on dozens of physical sensors on participating patients. Together, I see the insights from this work will lead to new, automated ways to measure aging in real-time. Once we can measure aging in real-time, we will be able to test new interventions and therapies at record speed.

Abbasi Lab is in the center of the action at UCSF. It offers incredible access to data and willing patients, and some of the most talented researchers and clinicians in the world. Our lab space is on the top floor of a gleaming new building, the UCSF Joan and Sanford I. Weill Neurosciences Building (pictured below).

Located in UCSF’s Mission Bay Campus in San Francisco, we are across the street from Chase Center and nestled among a number of other UCSF research buildings and facilities. Good coffee, food, and outdoor space are abundant, with easy access to public transportation. Biotechs and VCs are literally steps away, making it easier for discoveries to get to market.

Reza Abbasi-Asl is the Principal Investigator and leader of Abbasi Lab. Reza’s talents are what makes this possible. He draws talented graduate students and postdocs. He has high expectations. He takes the time to explore, discuss, debate, and coach. He exudes the contagious energy of a lab and researcher on the upswing. We’re developing a fun, synergistic culture working on a range of interesting topics. We expect to recruit many new talented, interdisciplinary researchers to join our efforts.

In the end, I’m here at Abbasi Lab because I see us building foundational tools for the fight to slow aging and offer humanity many more healthy years.

References:

[1] Marx, V. Method of the Year: spatially resolved transcriptomics. Nat Methods 18, 9–14 (2021). https://doi.org/10.1038/s41592-020-01033-y

[2] Jumper, J., Evans, R., Pritzel, A. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). https://doi.org/10.1038/s41586-021-03819-2

[3] B. Yu and Karl Kumbier, “Veridical data science,” Proceedings of the National Academy of Sciences, 117 (8) 3920-3929, Feb. 2020, doi: https://doi.org/10.1073/pnas.1901326117

[4] W. J. Murdoch, C. Singh, K. Kumbier, R. Abbasi-Asl, and B. Yu, “Definitions, methods, and applications in interpretable machine learning,” Proc. Natl. Acad. Sci. U. S. A., vol. 116, no. 44, 2019, doi: 10.1073/pnas.1900654116