#134 Alex Zhavoronkov: Unraveling AI's Crucial Role in Pharma

38:53So saving a life in our current context, in the traditional medicine context, is basically not letting the person die at this moment. So you might add a few years to life, but you haven't really saved it. And that was the paradigm that I was not very comfortable with. I'm still not comfortable with that. So when somebody says, well, this doctor saved my life, well, kind of yes or no, because you're going to die anyway. And not that you're not going to have significantly more years just because you were treated. You are still on the same path or maybe you lost a few years anyway to your life expectancy.

39:43And I was very concerned about this paradigm. So unfortunately, when I was growing up, we did not have the resources to go right away into aging. So I had to go into computer science, made some money, and then decided to educate myself in biotech. So did an entire nine yards in grad school and published a large number of papers. I also realized that aging is a much more complex biological process than most people think. So just looking at animals that live longer or shorter lives and we can either learn from them or extend their life and learn from them, That's not going to help much because we're just very different and we cannot become whales quickly.

40:42And even if we do, we marginally increase lifespan and we still accumulate many, many types of damage that we're accumulating today. So we are already a very perfect organism designed for long life. and I realized that there are many, many, many processes that you need to understand in order to understand aging. And I actually worked for a year for a company that was organizing many, many, many biotechnology conferences. It was called GTC Bio. I don't think they're around anymore, but they were one of the really top, top ones back in 2005 to 2010. and I helped organize over 30 biotechnology conferences in different fields, you know, modern drug discovery and development, massive conference, stem cells, regenerative medicine, vaccines, many, many, many, many others, and got to speak with many key opinion leaders because, you know, when you're organizing conference, even though you're still a young scientist, you can still get in touch with the top people and realize that aging is probably better to tackle it using AI, right?

42:04Because you need to integrate massive amounts of information. You need to, so one single silver bullet is not going to work. So even if you are going after a drug that kills aged cells, senescent cells, you would need to have the entire plethora of technologies to support that drug. You would need to first ensure that those pre-senescent cells are safe, right, and preferably rescued. After you kill those senescent cells, you need to ensure that there is no fibrosis. You need to ensure that those cells are being repopulated, those niches that freed up. They need to be revived, rejuvenated. You need to ensure that new cells come in.

42:54You need to ensure that the extracellular matrix is ready for that. Usually it's already stiff and calcified, mineralized. There are also advanced glycation end products in there and many, many other forms of extracellular junk. so it's very difficult for stem cells to go there. You need to ensure that your immune system works well and clears this junk at the extracellular and cellular level. You need to even account for systems-level damage and aging because very often, once you start an anti-aging therapy like the one I'm describing, and that's one out of hundreds of alternatives, you already have a lot of fibrosis in the tissue.

43:44A lot of really useful tissue has been replaced by fibrotic buildup, and you need to figure out how to clear that. Well, that's one of the reasons why my program one is going after fibrosis, because we realized that if you have a lot of fibrosis, that's it, right? We actually cannot do much so we need to figure out how to get rid of fibrosis and if you have an effective tool to do that great now you have a major uh you know element in your toolkit that you need to go after aging yeah um and and so you're attacking it uh sort of by identifying things like fibrosis that contribute to aging or contribute to disease as you age.

44:40There are, there's other work, and I'm wondering whether you've, you're involved at all. I had Vittorio Sebastiano on the podcast a year or so ago talking about his work with and Aca factors to reprogram cells and in effect roll back the age of cells. Have you done any work with that? So, of course, chemical reprogramming is something that we can actually do in the lab. So study that at scale, because now we can test hundreds of thousands of compounds on many different cell types and tissue types and different kind of organoids to see what kind of transcriptional methylation changes and phenotypic changes those compounds induce.

45:40and that's exactly what you want to study if you want to understand the potential of those drugs and drug combos for in vitro or in vivo reprogramming. So yes, we do that, but that's not our core line of work. It's a bonus.

46:04There's a whole philosophical, ethical debate about tackling aging because the population has doubled in my lifetime. I'm getting old and I'd like, you know, another 50 years if I could manage it. But the planet can only support so many people and the increased population leads to all sorts of other problems, you know, mental illness and, you know, conflict and that sort of thing. Have you thought about that? Or do you feel that those problems can be tackled once aging is not solved, but at least extended? Well, that problem is much, much easier to tackle. Yeah. So that's why very often people who do not understand science very well, they like to, or they're just not really good aging scientists.

47:19They try to go into those philosophical aspects, right? Because it's easy to debate. They can look smart and make all kinds of arguments. But it's very clear that the population is going to grow regardless of whether we are going to identify anti-aging interventions or not, because it just did. So if you are concerned about overpopulation, you should stop reproducing right now. Are you willing to do that? Probably not. Many people are willing to do that. So for example, I do not have children and I do not plan to in the near future just because I think I can be very productive as an individual without the need to pass my genetic material onwards.

48:15And usually we see that a lot of smart people do not really have successful kids. So it's better to support somebody else rather than to follow the evolutionary call and foster your own. So I actually like to give grants to students that demonstrate superior abilities. I don't think that my genes will make a difference. And many smart people do not have many kids. usually the reproduction is very correlated with intelligence and education and we see that in pretty much every country and every culture so once people get to a certain level of income and education they start having fewer kids and currently the population is still growing.

49:18It's predominantly growing in the developing countries where the level of education is lower and poverty is higher. And we cannot really do anything about it. The problem that we are facing are the people who have aged and are not doing productive work, but at the same time, they are drawing massive resources in terms of healthcare and social security. And they also vote. So in democratic countries, now the percentage of elderly is very, very high. And usually, they vote for something that they need. And we've seen what happened in France recently. The retirement age goes from 62 to 64 while the lifespan has increased to you know 85 plus it just doesn't compute right it's not economically feasible so the country will be deeply in debt and it's already not as productive as some developing countries and And people will still vote for politicians that allow them to enjoy early retirement and massive benefits.

50:43And we're going to see the second row. So we're likely to see the collapse of some of those civilizations that rely on the elderly to have the same vote as the young people. and we should probably worry about that, right? And not about overpopulation or how we're going to allocate the resources because otherwise we should just limit lifespan, right? Why not say that, look, the maximum lifespan that you are allowed to achieve is 75 done, right? So we are going to support you till that time extremely well. And after that, well, Let's not draw the Medicare, Medicaid, right? So I think people who are not ready to do that, they should not philosophize, right?

51:38Can you take us through the process, for example, with this IPF drug that you say is finished phase one trials? so just very relatively high level but not that high level because the listeners are mostly AI practitioners of some level so you you you pick a disease like IPF there are many targets presumably and then sort of starting there, what's the process? How do you decide on which target you're going to design a molecule for? So, sure. Well, first of all, I recommend the listeners to look at our papers that describe pandionics and also dual-purpose therapeutics. So if you just look for a dual purpose aging and disease therapeutics, you'll find a bunch of papers that was published describing the many methods that were used to identify promising targets implicated in fibrosis and other diseases.

52:53and also we just published a really fun transformer model called Precious 1 GPT which learns the basic biology from methylation and transcriptomic data and then transfers the weights onto the model which is trained on different diseases to identify targets that can be prioritized for age-associated diseases. And then we give it to Pondaomics to assess commercial tractability, novelty, and confidence using many, many different models. So this is pretty much the process that we went through with our program one with the antifibrotic. So we looked at the targets that are likely to work in basic aging and also in a variety of fibrotic diseases using different machine learning techniques.

53:48then we prioritize 20 targets that are less heterogeneous in the patient population because you want to see something that is occurring in many many many people and scoring at the very top it might not be the number one or number two target but it's going to be consistently high in many, many people. So that's what you want in order to achieve maximum benefit for everybody and for as many as possible. And then we prioritize 20 like that. We managed to test only five because for many of the top 20, the assays were not available. So that's actually one of the big problems that you face in biological AI, right?

54:40So very often you come up with a really great hypothesis that's validated by other pockets of evidence. But in order to perform the experiment, you really need to be able to have this biological target available as a protein structure, as a protein, as a ready-made protein. and you need to have a molecule to test it or you need to be able to knock out that gene and very quickly see if it works in a specific disease model. And very often it's just not available. So we could test only five. And one target showed, so all of them worked. But one target was scoring at the top. And also, we realized that drugability is a big feature, right?

55:33So if you go after the top scoring target, but it's not druggable, you might never be able to deliver a drug, right? On the market, you run out of money before you demonstrate that you can drug it. What does that mean, drugability? So it means that some proteins are easier to develop small molecules or antibodies against than the others. So there are grand prizes in chemistry right now, like, for example, CIMIC, a very famous oncogene. So if you can drug it with a small molecule that is safe and effective, and you can disable CIMIC or some associated proteins in the complex, it's a grand prize.

56:23It's billions of dollars, right, because many cancers are driven by CEMIC. However, it's very difficult to drug. So we have spent several years developing a degrader now for CEMIC, right? And then you realize, okay, even the biology of this target is very difficult, right? It's just being produced very, very quickly, right, and degraded quickly. So if you figure out how to degrade it, you also need to ensure that you degrade it extremely quickly, quicker than it gets produced. So there are many, many challenges with some of those difficult to drug targets because sometimes it's just flat surface.

57:12There is, you know, imagine a wall that you want to climb on and there is nothing to grab, right? So some of the proteins are like that. And you want to ensure that the complexity of the protein and the... So you need to find the target that you can drug within a certain time, right? Before you run out of money and without a huge amount of resources. So some targets, they have a lot of those. So some of those mountains, they have a lot of those pockets that you can grab. And nobody has grabbed those pockets before, but you see that they are there. And it's better to go after targets that have a good balance of confidence and commercial tractability and novelty and also drugability.

58:13yeah so if you know that it's going to be extremely difficult to drug you are increasing the risk of the program yeah and and you're you're talking about

58:28proteins and and the drug ability depends on the shape of the protein is that right that's one of the features yeah so it's not only the shape sometimes it's it's part of a very large complex that is not exactly well understood sometimes it is the pocket that you need to drug that is part of the protein protein interaction uh and that's very difficult to uh to drug and to fit the molecule in that specific pocket right um so it's the structure but also many other things yeah and also going back to uh targets for which you don't have assays i didn't i didn't understand that what what does that mean i mean there are is is that proteins that for which they don't know the shape or what what is the uh issue so uh now you can actually do um something similar to alpha fold to create the shape right but it's just very often it's very difficult to drug even if you have the shape.

59:41So again, it's either flat surface or it is the structure for which you know that a lot of groups tried over the many years and failed to deliver. So usually you don't want to go that route unless you are very, very confident in your system. And at that time when we started program one, we didn't know that we could drug. many other protein classes. Now it's an industrial strength platform and we have much higher confidence in our AI that manages to achieve very high levels of drugability. But at that time, we decided to go after a very balanced target. So we knew that we can drug it. It belonged to the protein class that is druggable.

1:00:35and it didn't have any molecules available for it, any good ones. So we decided to go after a program for that specific target. Yeah. And how long, for example, with this IPF drug, from identifying that you want to tackle this, to identifying the target, to identifying, to generating a molecule or multiple molecules and then evaluating, deciding which ones you want to put into preclinical trials and then clinical. How long is that process now using your system as opposed to what it has been historically? Yeah, so for our program one, to get to preclinical candidates, stage, preclinical candidate stage, meaning that we have validated the target.

1:01:37We have already selected the molecule out of the many options, out of those many perfect needles that we generated, we selected one. And we've conducted substantial safety studies on that perfect needle and on the target. So that's preclinical candidate. So from zero to preclinical candidate, it took us 18 months originally, which is actually for a new target, it's a very short period of time when you're talking about really good molecules that are also easy to synthesize. And then it took us 30 months to get into phase one, and phase one took almost a year to complete. So it still sounds like a very long time because it's like 40 months, right, to phase one complete.

1:02:36But so just under four years. But using traditional approaches, it takes significantly longer. So usually it would take you at least two, three years longer. And the probability of success, probability of getting there is much, much lower. so now we have done that many many times over so now we've nominated 13 preclinical candidates to date right so nine preclinical candidates just last year and to put it on put you in perspective big pharmaceutical companies using internal rnd uh in small molecule space they usually nominate four to five yeah uh and i've spoken to daphne collar uh about her company i've had her on the podcast and she's uh if i understand it was a number of years ago but but she's looking at the uh at the the universe of molecules and trying to identify based on uh the qualities of of of of similar molecules, a space in which to search.

1:03:49Am I getting that right for then choosing existing molecules to test? Is that right? Is that the way that the research has progressed? I mean, she's using AI, obviously, to speed it up. So, you know, everybody is using AI nowadays. High school students use AI. And I don't know what Daphne Keller is doing, right? So as a matter of fact, you know, the level of transparency there is not exactly very high. They like to provide really great PowerPoints, but we really want to see the results. So I guess once we see her molecules in phase one, phase two clinical trials, then we probably and also some kind of record of how they were discovered.

1:04:48And then we might be able to do some comparison. But currently, I just have not seen any of that. Right. So it's the technology is being developed, I guess. So it's a pretty credible scientist, but I'm yet to see the value of that research. So currently, there are only several companies that managed to take AI-designed or AI-discovered molecules into human clinical trials, and there is less than a handful of them. So I don't think Daphne Collar's company is in that league yet. Yeah. Your process that you've created, presumably you can begin doing it, you know, scaling it up, doing it in parallel or, you know, licensing it to others to do it.

1:05:55Do you think that we're on the cusp of a huge number of new drugs and of solving or curing or at least treating many, many diseases that have been untreatable? So yes and no. So yes, we can definitely generate a very large number of preclinical candidates and even clinical candidates very quickly. So we've demonstrated this ability. And in order for us to significantly progress, the pharmaceutical companies need to change a little bit as well. Because the level of productivity within the pharmaceutical companies has been historically low. one curse that we see within the pharmaceutical industry is that the management changes very quickly right so usually so I'm in silico since 2014 I within pretty much every pharma company with the exception of you know maybe a couple I've seen management changes three times right So they basically live on three to four year cycles and then they recycle everything.

1:07:13And the level of productivity in internal R &D is very, very low. So yes, there are sometimes very promising and really great results. So sometimes they do produce bloodbusters, but most of the time they don't. And the model has to change so that the pharmaceutical companies focus more on licensing and very rapid progression of the programs into the proof of concept. So phase two complete, right, in the areas where you've got significant unmet medical need. and the partnering needs to just speed up, right? So I actually think that nowadays they should be using AI to pick the partners because many humans are just not effective, right?

1:08:15I mean, chat GPT can actually very rapidly show you, okay, which company has which target, which is better, which has demonstrated more results, which published more credible work. For big pharmaceutical companies, what we see is that a lot depends on relationship and trust. So they usually build up relationships with scientists and other biotechnology company executives that do the same thing all over again. and they just partner within a very small circle. So I think that there needs to be a very effective partnering strategy within at least some pharmaceutical companies based on merit and based on equal opportunity and based on the good to the patient.

1:09:16So how do we benefit the patient the most? and of course assessing commercial tractability. So once this happens, we're going to see massive acceleration. But in order for us to get there, we will see a lot of failures, right, in between. So we see that some pharma companies like to partner with university startups that don't have anything, no proven concept, nothing, right? So they support them with 20, 25, 30 million upfront. And then they hope that something is going to happen. And historically, it just never happens. Or you see the bets after phase two complete. And those will be in billions of dollars, right?

1:10:08So many billions of dollars spent. And in the meantime, they could have made 100 bets. with lower money, right? And much cheaper and tried many more approaches, right? So instead of trying to get one de-risked asset in phase two, they can make 50 bets on something that is a little bit more risky in terms of commercial attractability, viability. But once the process is streamlined, it could dramatically impact R &D. But it also needs to start with investors, right? And I'm talking about large institutional funds that currently would not invest in pharma companies that don't spend, let's say, 15%, 20 % of their revenue on internal R &D, right?

1:11:05As just standard. And they just have to put the money in. And the amount of money is massive. We're talking about for Novartis is$9 billion a year. Right. So, of course, a large part of that are clinical trials. But much of that is basic research. Right. And well, it's good. It's great. Right. But where are the drugs? Right. And one idea is, you know, those$9 billion could be spent on partnering with companies like ours at a little bit earlier level and just focusing on clinical study design, on ensuring that the drug reaches the right patient, on ensuring that the drug is combined with the current drug within Big Pharma that has demonstrated efficacy, and that is improving patients' lives, and working with the FDA to ensure that the outcome measures are consistent with what the FDA wants, because FDA is the ultimate authority.

1:12:14By the way, a lot of people criticize the FDA. I can tell you that in some areas, this machine is a Ferrari, right so they have already so that those many many patients and patient advocacy groups ensure that they have standardized response times right you know 30 days 60 days 90 days so you have very very quick interactions with them they're very open so I think that the integration on the clinical level and also regulatory level that's something that big pharma companies should focus on more, and then they just need to streamline and automate the partnering apparatus and introduce AI there. So that for all the AI folks that are listening, if you create a really nice partnering portal that matches the demands of big pharma with the supply of high quality assets, that's what's needed, right?

1:13:18You just need to be able to remove the huge overhead inefficiency and waste in this partnering space. So then it's going to be dramatic. Of course, we're also working in this space. And you know me, right? So if I say something, I make every effort to deliver. so we are on the track to deliver those partnering deals. And we've partnered with many, but we, of course, want to just streamline that and ensure that it's as automated as possible. Yeah. Okay. Well, we're over an hour, but the last question on the longevity or anti-aging, um uh i mean certainly lifespans have have increased and in our life times uh but uh do you think uh do you do you expect uh results that will impact your life and uh how long do you think you'll you'll live with with those advances so you know to answer this question we developed an annual conference.

1:14:40It's called ARDD, agingpharma.org. It's the largest one in the world in the field of aging biotech. You know, hundreds of key opinion leaders travel there every year. So end of August, beginning of September, five-day event. So, you know, you're going to see David Sinclair, Vadim Gladeshev, Eric Verdon, so all the big names. They present their research and then we try to assess, you know, how far did we process for progress. There is now also the longevity biotechnology association. So with that kind of disclaimer, I think that we might be already there. Right. So we might be already on the path of longevity escape velocity where, you know, the longer we live, the longer we should expect to live.

1:15:33because technology is becoming exponential. Unfortunately, there are many threats to that, right? Because first of all, it's not for everybody. So if you're 70 right now, that exponential touch might not be soon enough for you, right? But it might be there for the current 20-year-olds and 30-year-olds, right? Because, again, the level of progress that we see in the world is huge, right? And now when I entered the field, there were no books. There were a couple books, three books maybe, that you could read on the biology of aging and the current state of affairs. Now even I wrote one, and it's already old.

1:16:32And now we have courses that thousands of physicians take and thousands of scientists. We now even have, you know, investing in longevity course for free, many hours by pharma experts, by VCs, by medical doctors, which is actually a cautionary tale, more or less. And currently, it's very hard to say where we are, but we are definitely moving closer to the longevity escape velocity. There are many factors like geopolitics, for example, that might impede this progress dramatically. For example, there are dictators with a red button. They can push right now and the entire world goes to hell, right?

1:17:23whatever progress we've achieved and we're probably as close to midnight as we have ever been. So a lot of people are thinking longevity bunkers to ensure that we at least can preserve some knowledge and some people. But in general, I think that we are on the exponential track and I don't like to make predictions in terms of how long you can live. You should not have that limit. If you want to have a limit, set something achievable, maybe 120, 150. 120 has been achieved already. So how much worse are you than Jean Calment? I don't know. A lot of people don't want to go that route, but intuitively, they do.

1:18:14They just don't want to make an effort. and I think that nobody should be should be pessimistic so optimism is the number one component in longevity yeah okay that's terrific actually let me ask one more question I may cut it but what is this multimodal transformer based aging clock that I've read about all right so precious one gpt oh that's precious one gpt yeah that's the precious one gpt so we've got methylation and uh transcriptomic data used for prediction of age so you build a massive transformer which is trained on many many many data sets of biological data and metadata to predict age and actually that's not going to be as accurate as a purpose-built fit forward deep neural net, for example, but it can make a very useful, uh, inference into the various biological processes that are not apparent.

1:19:23And then you can go, you cannot get out of, uh, other deep learning techniques. And then you transfer the weights, uh, onto the model, which strain on diseases. I see. And, and you're, you're looking at, um, you're, you're predicting, uh, I don't know. Just to explain that, you're predicting the age of a person? Yeah, so originally you built a predictor, so that system can do many tasks, right? So it can perform many tasks. One of those tasks is prediction of chronological age. So basically train the transformer to predict the chronological age of a person using two data inputs, methylation and transcriptomics, and of course, a lot of metadata.

1:20:20So this transformer learns the fundamental basic biology of aging, right? And you can also query it for the different relationships between methylation sites and expression of individual genes. And it can do many other things, right? But it can predict age. Now we are working on another release which will help you generate biological data, so synthetic methylation and transcriptomic data with age as a generation condition and a few other things. So those transformers, they acquire the ability to, again, comprehend the basic biology. and then you learn from basic biology into diseases aging research is the most impactful area you can go into you can go into ai safety you can go into whatever security but everything is valuable but aging research will produce the maximum number of quality adjusted life years for everybody on the planet you develop a drug that increases everybody's life by one year, you just generated 8 billion life years in today's terms, right?

1:21:45If you look at net present value of people are producing, you generated many, many billions of life years. So that's a very worthy goal to go after. And before you go, give Shopify a try. It's only a dollar a month for their trial period, and you can sign up at shopify.com slash ionai. That's shopify.com slash ionai, E-Y-E-O-N-A-I, all run together, all lowercase. Go to shopify.com slash ionai to take your business to the next level today. Shopify.com slash E-Y-E-O-N-A-I That's it for this episode. I want to thank Alex for his time and wishing Godspeed in developing a cure for aging, a subject that is ever more urgent for people my age.

1:22:44If you want to read a transcript of this episode, if you want to read a transcript of this episode, You can find one on our website, Eye on AI. That's E-Y-E hyphen O-N dot A-I. In the meantime, remember, the singularity may not be near, but AI is about to change your world. So pay attention.