In this session, Alex Zhavoronkov, CEO and Founder of Insilico Medicine and Deep Longevity, offered his roadmap for what he calls Longevity as a Service. He explained the approach Deep Longevity and are taking, the various clocks and machine learning methods they are using, the necessity for physician education and coordinated building of a longevity ecosystem spanning physicians, clinics, insurers, academia, pharma, and more. And he also went into detail about a promising new interesting field of longevity psychology and connected subjective and psychological clocks.


Presentation: Alex Zhavoronkov

  • Deep Longevity is a spinoff of Insilico Medicine, it was spun out in 2019 and acquired by Regent Pacific (HK: 0575), founded by Jim Mellon, now focusing on longevity and health in China.
  • Register for the annual meeting of the Aging Research and Drug Discovery community – ARDD2021, sponsored by the major journals.
  • This year we’re going to have a Longevity Medicine Worksop to focus not only on the science, but also on clinical applications. We cannot do Longevity as a Service without the physician’s buy-in, so we’re doing a lot of work in that.
  • Part of that work is the Introduction to Longevity Medicine for Physicians course on Udemy (and free on Longevity.Degree). It has ACCME accreditation, so if you are an MD you get 2.7 AMA credits and a medical certificate. We are making a portal for longevity education now, because physicians are the core bottleneck and central driving force, that’s why we cannot ignore them and need to include them. Our industry will move at the same speed as the physicians.
  • The field has matured from when Aubrey started the movement, but people are still afraid of giving themselves false hopes.
  • How can we provide longevity as a service today? The toolkit is very limited right now, and if we are moving the same seed as physicians, chances are it’s gonna be a little too late for baby boomers for example. That’s why we are trying to define the field of longevity medicine, which is a cutting edge research, science, therapy and diagnostics, currently primarily focused on participatory medicine and personalized science, powered by biomarkers of aging and longevity, predictive and prognostic, data-driven, personalized and preventative, all that with the help of artificial intelligence.
  • In the context of traditional preventative medicine, the doctor looks at specific age ranges (like 60-70) and compares their well-being to the people in the same age range. So if you don’t have too many comorbidities in your 60s, you are still considered healthy, however if you would compare yourself in the context of your entire lifespan, then of course you’re much worse off.
  • And what we want to do is to look at the whole lifespan and bring people to the optimal healthy state (like 30 or 40). The question is how to do that, and how to track that (and repair and rejuvenate – basically how do we apply SENS into action).
  • First we need to educate and involve, that’s why we made the course.
  • Then you need to do advanced anamnesis and analysis based on data – collecting data over time.
  • Comprehensive longitudinal diagnostics – we need to collect data over time. If you’re not collecting data, you’re losing, because you don’t have the benchmark of optimal ranges for yourself.
  • We need to identify longevity bottlenecks – the clocks that are ticking faster for a specific person and drive their aging in other tissues and organs as well. For example in the general population we can see lung, liver, kidneys, bladder ticking faster than any other tissue types, like muscle. However in the context of an individual, everybody can have different bottlenecks.
  • Then we can look at personalized risk & benefit analysis of interventions. Basically use the same view for longevity products as for financial products – some people are more risk-averse than others, and might be open to less or more proven interventions. So the individual and their longevity advisor may figure out specific guidelines based on their risk profile. We are now coming up with this risk profile we call “know your patient” strategy and questionnaire to understand that risk and educational profile.
  • And then when we slow aging with different interventions, we need to measure the pace of that with aging clocks that measure every level of human organization.
  • Doctors won’t just engage without clinical protocols and proper training. A lot of stakeholders need to engage in order for the doctors to be able to follow.
  • First is academia, we see major advances in model organisms. New companies and geroprotectors, new academics. However, for the academic breakthroughs to come into the pharmaceutical clinical trials, it usually takes 5-10 years, and the clinical trials take another 5-10 years. For example the technology for checkpoint inhibitors for cancers, it took like 20 years for the products that are on the market now to get there from the academia. So academic science needs to be supported, but we also need to look at what can be done today.
  • In the pharmaceutical industry, the efforts are not really prioritizing the R&D for clinical use, they need to spend some amount on R&D. They try to be prepared for what’s coming in the biotech sector, but it’s not their main motive. So we should not really expect much progress from the pharmaceutical industry within the next 10 years. We should engage them seriously but we shouldn’t expect too much, they are driven by the market, and currently nobody has yet put longevity medicine and aging by itself into phase 3 trials.
  • We should not expect significant progress from the clinics as well. They are driven by medical doctors and also by scale – to institutionalize an intervention and make it big, there would need to be some mainstream longevity clinics at scale. Some are starting in the US, but all of them lack scale.
  • My big bet is on the insurance companies. They need to innovate and they are trying to look for additional tools and innovative products for both customer acquisition and underwriting. But there is no major central effort. So we need to figure out how to turn this into a longevity ecosystem faster. And that’s why we need physicians, so they can be the driving force to demand it.
  • Right now, we’re still operating with basic recommendations that you could get from your mother, like diet, sleep, exercise. So the near term outlook is pretty bleak. But the near future seems better. There are some things in the pipeline close to or already on the market, but these are things that the physicians can already play with.
  • What is coming up are multiple interventions that can be combined.
  • The first thing we can do is to measure aging. Biomarkers of aging are now becoming more available and popular. But we are really at the beginning. We still don’t know much. We don’t understand exactly what the biomarkers are telling us, how to properly combine them, interpret them, and improve them. But we can already predict age and start playing with tweaking it. We’re trying to get to the point where we have a digital twin, but we’re still really far off. That’s why I really appreciate the work of scientists like Vadim Gladyshev who are trying to dig down into the very basic biology of aging and then going back and seeing how we age via the biomarker lens.
  • Methylation clocks are the most precise but also the most difficult to interpret or carry into a therapeutic intervention unless you’re doing epigenetic reprogramming.
  • We use GANs a lot to create synthetic data (like the images you see) and also for biomarker discovery, target identification. We can give the network some data and then ask the network to generate us data based on the patterns in the original data. We can ask for data throughout the time from current age all the way to let’s say 120 and see what changes to identify new biomarkers. So similarly to aging a picture, we’re artificially aging a data type and looking at what changes. That’s how you can also identify targets and promising interventions. The same techniques can be used in many different areas of human biology and chemistry research. We also use it for molecular generation (that’s what Insilico Medicine is doing), and for many more activities.
  • So even from synthetic data we can now derive the most valuable biological data which is age.
  • The way we do it is with feed forward neural networks, convolutional neural networks, and other machine learning techniques. We feed specific data type into a model annotated by age with as many features as we want, and we can try to go and understand basic biology and even psychology from those networks, so we’re now experimenting a lot with it. Even with uncommon data like psychological questionnaires. And you can also try to understand causality of the different biological processes by predicting age in different age groups and looking how important the features are for the prediction.
  • Then you can also add another neuron for health status like a disease for example, we’ve done that recently at Insilico. We basically train on age, then retrain it on different types of fibrosis, identify the valuable targets by looking at the different features that might be causal and then we put them into pathways, establish biological relevance, and identify targets. You can do this for many other processes, and it gives you a pretty good business model, because if you can identify a target for age-related disease, that provides you with a clinical pathway where you can do a net present value on an asset and a pretty substantial valuation for your company and fundraise for additional research.
  • There are multiple biomarkers of aging that are using deep learning. We published the first blood based clock with a reasonable mean absolute error of 5-6 years, which is now a workhorse clock for us since 2016. We’ve published and sometimes patented many different clocks.
  • Here are our efforts in this area.
  • We’ve developed a variety of aging clocks to cater to different audiences and use cases. Most of our work is in deep hematological aging clocks, where we can now through generative approach take a very small number of markers and reconstruct the rest using GANs, if we are lucky. But the more markers we get the more valuable the prediction is and the more utility it has, because we can better see which features we can change to look younger to the network. But of course the most valuable clocks are the deep transcriptomic and proteomic clocks, there you can derive valuable targets from the clock, but they are also much more interpretable, but at the same time variable.
  • The very simple idea or the concept for a deep aging clock is depicted here. Take a very large number of features, let’s say 50 markers, and a very large number of profiles annotated with age, and train neural networks with the features of for example blood markers in your input layer. Then you add many interconnected layers of neurons and one neuron at the end predicting the biological age. You can use one network or stack them in an ensemble.
  • This gives you a pretty reasonable accuracy, in our first clock we got 5.5 MAE, and at that time we started noticing that people who are older than their chronological age have more comorbidities and are less healthier and therefore substantially increased hazard ratio, and vice versa.
  • We’ve also demonstrated that in another paper where we compared populations in different countries and found out that if you have your biological age older by 5 or more years compared to your chronological age, your hazard ratio increases substantially. And vice versa. This is extremely important for underwriting for insurance companies. And also for clinics, to explain and tell a story why the clients should care about their results, because they are tied to morbidity and if their clock is bad, it increases their risk of dying from any cause. This makes people on each end more engaged, which is an essential step for connecting clinics with insurers – this is how we can provide Longevity as a Service in the future.
  • We have published multiple clocks with the same philosophy but different data types.
  • We developed two tools that are trying to bring the ecosystem together. One is which is trying to put multiple aging clocks and data together. It will tell you which features are moving the predictions the most and then you can work with your physician to try tweak it and lower them. We also work with 18 premium clinics that are experimenting with the clocks (like Human Longevity in San Diego, Hooke in London, Life Hub in Hong Kong, Boulder Longevity, and many others). So now we are getting quite a lot of anonymized data for before and after interventions. The networks provide age metric reports for those clinics and patients, and also recommendations based on their metric. 
  • The workhorse algorithm is optimized for 39 markers, but 31 is also possible. Those basic markers can tell a really interesting story about different organs.
  • The process is that you take a blood test, request a report, review the report, make changes, and then follow up. We are getting some interesting data from the clinics about different interventions, so we hope to publish it soon.
  • Those age metric reports are also interpretable, some more (psychological for example), some less, so people are trained to tell the story, and provided with tips on what to change to “get younger”.
  • You can also identify which biomarkers contribute to you being predicted as being older, and you see which features to tweak to be younger according to these clocks.
  • You can use all kinds of interesting tools like self organizing maps to see where you need to be in terms of your parameters to be in your optimal healthy state and also to your 20-40 biological age.
  • One other important aspect that we try to investigate right now is psychological age. Can we predict how old you feel subjectively and psychologically, so we can as humans interpret those features?
  • We came up with a hypothesis that biological and subjective ages are connected, and we looked through multiple longitudinal databases with surveys and biological data, and connected them to multiple other data sets.
  • From that we came up with two important features – psychological and subjective age.
  • We are looking at a very large number of modifiable factors (excluding non-modifiable), predicting age from them and correlating that with other mental health factors.
  • With PsychoAge we’re now around 6 years mean absolute error and SubjAge about 7 years mean absolute error, which is pretty good, I didn’t expect that.
  • We demonstrated that both psychological and subjective age correlate with mortality, so we can now predict it the same way we do with BloodAge. Subjective age seems to be a very good predictor of when you are going to die.
  • Then we deconvoluted the networks back into different risk factor categories to see which ones contribute the most to the predictiveness.
  • The most impactful was longevity expectation, and the best way to change that is to systematically deceive yourself that you will be living longer, and then positively reinforce those beliefs further and further. Optimistic attitude towards health and longevity moves the longevity expectation into the right direction and makes you psychologically and subjectively younger, and is also the easiest to implement – by stretching your longevity expectations. So if you wanna take one thing from this lecture, convince yourself that you will be living longer, because the conviction itself contributes to longevity.
  • We are also starting to link in data from wearables.
  • This is the tool that we envision to drive Longevity as a Service. We have some pilots with insurers, clinics, diagnostic centers, pharmas, beauty clinics, wellness centers,…
  • Major stakeholders are clinics, insurers and employers.
  • One of the best partners is Human Longevity clinic, which provides wealthy clients with very comprehensive plans and uses our clocks, which means a lot of data for us.
  • Now we are trying to build the longevity network by partnering with clinics and insurers. That’s why it is incredibly important to standardize the clocks.
  • There is also a hospital path. We want to explore all kinds of experimental interventions and turn longevity products into essentially venture capital products where you create an actuarial model and look at how much life you gain and how much risk you accumulate in the context of your personal longevity.
  • Checkout out the education program, where you can learn more about mentioned concepts.
  • Register for this conference, this year on site in Copenhagen.
  • Thank you very much and let’s connect on WeChat.


Has YoungAI had any discussions of partnering with Forward Health?

  • Not yet, but they are as close to Longevity as a Service as you can get – a sort of McDonalds for diagnostics as I think it should be. So we would be happy to partner, please connect us if you know them.


Linear regression works really well with epigenetic clocks, almost seems to be too good to be true. But my question is about deep learning and the issue of having interpretable models. In your experience what is the breakdown between where linear regression does the trick and then when linear regression is not enough but you were able to make it work with deep learning.

  • You can look at this paper and this specific figure we did where we compared different algorithms. The important thing is to see which features were prioritized in different architectures, so you can then combine a few different algorithms for where they work the best. We also learned from that work the value in combining and comparing different algorithms in different tissues.


Can you clarify how you feel about the usefulness of epigenetic clocks? You said that it is not the most actionable and interpretable – do you mean that it does not respond quick enough?

  • It’s the most accurate clock we have today, but there is always a tradeoff between accuracy and sensitivity to interventions. Some clocks are extremely variable, where one person can be younger and older on the same day. But that’s exactly what you want – you want to understand what makes you younger or older on the same day as well. But also what is it that makes you older or younger over longer periods of time.


I think it is a bit naive to reset some blood markers to younger age, like cholesterol where for example middle aged women with higher cholesterol are more protected with a bit higher cholesterol.

  • Yeah I agree, we don’t know the complete story with most of the things. We’re just showing what needs to be changed in blood for you to be predicted younger by the deep learning net. Think of it as wrinkles with photogenic age – you can use for photo face age, we did a paper together where we demonstrated that you can build an age predictor from face photos that is 2.2 years accurate. And then you can use different techniques like covering different parts of the face to see where the major features that make you older or younger are. That study has shown that the area is around the corners of the eyes, so if you fix the wrinkles there, you are predicted to be younger by the neural net. Of course that doesn’t necessarily mean that you’re going to be healthier or younger, but you are going to be predicted younger by the network. We follow the same kind of philosophy with blood and other markers. I’m not saying that it is right or wrong, it might quite possibly be wrong. But we can at least show that you could be predicted younger if you change some features. I cannot say that it will necessarily make you live younger yet though, for that we need longitudinal data and nobody has done that before. And if we don’t study this area in humans, we will never know.


Are you currently partnering in any fields beyond pharma & cosmetics and/or are you planning any other cross-disciplinary applications in the future?

  • Constantly thinking about it, I am very interested in partnering with anything tech. Ideal partnering would be with Amazon, because they could provide Longevity as a Service in a very true nature and the true meaning of this word. Because they help you waste a lot of time now with Amazon Prime Video, so I think it is only a matter of time before they realize that in line with ESGs, they should return some of that time back through longevity. And they collect a lot of data about people. Partnering with organizations that waste a lot of time to return some of that time would be a dream. Currently they lack that vision but I hope that I can ignite somebody to spread this longevity virus to them as well.


What are the next steps for the individual services and what can this group help with?

  • We have the Deep Longevity app for which we are hiring. For chief positions in business development (if you have experience selling to insurers or clinics, let us know). But also for many other positions if you’re excellent at what you do. Or connect us with players that are in those spaces.
  • Insurers will not change without medicine, medicine will not change without insurers, pharma companies will not change without medicine and insurers, academia is not going to accelerate without a huge consumer business. In order for us to drive Longevity as a Service vision, we need to ensure that all those areas are progressing very quickly and at the right pace, and we cannot progress faster than physicians and medicine, so that’s why we need to ensure that medical doctors are engaged, involved, and educated.


Do you have any predictions for machine learning or aging clocks? Interesting challenges to take on?

  • The best way is to join one of the companies or academic groups that are already working in this area. I think there are many more data types that we could work with to predict age and try to understand it better. If you know just ML and don’t know much biology, go into imaging – MRI, CTs, or especially ultrasound, where there is a lot of opportunity for progress. Try to build a clock based on such an accessible data type and get it into the clinic. If I had more time, I would go to ultrasound because it is a way to get non-invasive diagnostics quickly. And then I’d try to combine the ultrasound data with something like gene expression, or mineralization levels, or cross-links. So the combination between accessible imaging biomarkers and some exotic biomarkers, to make them cheaper is the way to go. If you can substitute some of them or get the same answer by using a cheaper method, that is extremely valuable.


Have you seen ultrasound to work well for making clocks in other species?

  • I don’t know, ultrasound is one of the data types where we have more from humans than animals, so not sure.


Any other challenges?

  • I would also go into monkey business. Looking at experimental monkeys, it turns out that most of the monkeys being supplied to pharma are like 3 years old. Because if they are below 3 years, they are immature, and over 3 years old are too expensive to handle, so most of the experiments are done in this age range, no-one is experimenting with them longitudinally. If we had more data about monkeys from that point of view, we could learn a lot. Mice just do not live long enough to accumulate the damage humans do and therefore we miss a lot of insight we could have from monkeys. Like for example the cholesterol question, seeing that in monkeys with the ability to take a tissue sample in a little bit easier way.


If I am now a potential customer, which site should I go to?

  • We’re B2B as the major segment, so from the consumer perspective, you can download the app and let us know about the UX, and report what you do and don’t like, that would be helpful. We’re trying to put the customer in the front. So you can track your blood age over time and opt out from sharing data anytime.
  • If you are running an insurance company or a clinic, please do come to us, we want to partner with you.



Seminar summary by Bolek Kerous.