Scroll Top

Insilico Brings Generative AI to Drug Development and Discovery

Episode Highlights

  • Drug development and discovery is the process of creating new medications or improving existing ones.
  • Preclinical animal testing is conducted to evaluate the safety and efficacy of the potential drug candidate.
  • Clinical trials in humans are then conducted in several phases to evaluate safety and efficacy further.
  • Regulatory approval from agencies like the FDA is sought to ensure the drug meets standards for safety and efficacy.
  • The patient is the most important stakeholder in drug development and discovery.
  • Relentless innovation is essential to keep pace with technological advancements and improve drug development processes.
  • Insilico Medicine, led by Alex Zhavoronkov and Dr. Ren, focuses on software development and drug discovery.
  • Chemistry42 is a generative system that optimizes identified targets by redesigning known molecules with desired characteristics. It is primarily used for lead optimization and can generate molecules with specific properties based on user-defined criteria.
  • Chemistry42 employs over 40 generative engines, including GANs, transformer neural networks, and genetic algorithms, to generate molecules. A predictive system evaluates the generated molecules using over 500 criteria and provides feedback to improve the generative models through reinforcement learning.

Full Transcript

The Harry Glorikian Show 

Insilico MedicineAlex Zhavoronkov, CEO 

Final Transcript 

Harry Glorikian: Hello. I’m Harry Glorikian, and this is The Harry Glorikian Show, where we explore how technology is changing everything we know about healthcare.  

It may feel like generative AI technology suddenly burst onto the scene over the last year or two, with the appearance of text-to-image models like Dall-E and Stable Diffusion, or chatbots like ChatGPT that can churn out astonishingly convincing text thanks to the power of large language models. 

But that’s not really true. As with most every technology revolution, the real work has been happening in the background, in small increments, for many years. 

One demonstration of that comes from Insilico Medicine, where my guest this week, Alex Zhavoronkov, is the co-CEO. Since at least 2016, Alex has been publishing papers about the power of a class of AI algorithms called generative adversarial networks or GANs to help with drug development and discovery. 

One of the main selling points for GANs in pharma research is that they can generate lots of possible designs for molecules that could carry out specified functions in the body, such as binding to a defective protein to stop it from working.  

Drug hunters still have to sort through all the possible molecules identified by GANs to see which ones will actually work in vitro or in vivo. But at least their pool of starting points can be bigger and possibly more specific. 

Alex comes from a background in computing rather than biology. And he says that when Insilico first started touting this approach back in the mid-2010s, few people in the drug business believed it would work. 

So to persuade investors and partners that the technology works, the company decided to eat its own dogfood, so to speak, and take a drug designed by its own algorithms all the way to clinical trials. It’s now done that. 

This February the FDA granted orphan drug designation to a small-molecule drug Insilico is testing as a treatment for a form of lung scarring called idiopathic pulmonary fibrosis. Both the target for the compound, and the design of the molecule itself, were generated by Insilico’s AI. 

And with the orphan drug designation, which supports the development of drugs for rare diseases, Insilico is now in line for potential federal grants, tax credits, and extended market exclusivity, assuming it can get all the way to FDA approval. 

The designation was a big milestone for the company and for the overall idea of using generative models in drug development and discovery. 

I asked Alex to come in and talk with me about how Insilico got to this point; why he thinks the company will survive the shakeout happening in the biotech industry right now; and how its suite of generative algorithms and other technologies such as robotic wet labs could change the way the pharmaceutical industry operates. Here’s our full conversation.  

Harry Glorikian: Alex, welcome to the show. 

Alex Zhavoronkov: Very happy to be with you. Harry. 

Harry Glorikian: So, you know, as I was just saying, it’s been a while since I’ve seen you. I remember when you were telling me about Insilico, you know, I feel like it was 2015, 2016, definitely way pre-, pre-, you know, COVID. But Insilico has so much going on in so many areas. I mean, I’m trying to frame this for the people that are that are listening but for example, I know you provide AI drug development and discovery tools for other companies, but you also have a pipeline of your own drugs that you’re developing. So I was hoping maybe that you could start by telling us how do you boil down a description of the company for people who are not, say, familiar with it, right? Is there a single technology or a theme that unites everything that you do? 

Drug discovery and development podcast - The Harry Glorikian Show

Alex Zhavoronkov: Sure. Well, again, it’s a great pleasure to be speaking with you again on record today after so many years because, yeah, 2015, 2016, it was even difficult to explain what generative AI was. And right now the consumerization of that technology made it very easy to explain. Right? Because you remember I was probably showing those flowers where you say, okay, add one more petal, here is a petal and we can do the same with molecules or synthetic biological data. And people did not get it. And many of them criticized this technology and this approach. And also some of them still do. And what unites Insilico currently are both the services and the internal pipeline is patient first. So we are very interested in delivering drugs to the patients as quickly as possible, either through software or collaborations or through our own internal pipeline. So about two years ago when I first met my co-CEO, well three years ago now, when I first met my co-CEO, Dr. Ren Feng or Feng Ren, depending on how you, where you are, PhD from Harvard, 11 years at GSK, and then he listed a company called Medicilon. It’s a big contract research organization which competes with Wuxi Apptec. And at that time it was, you know, $8 billion. Once you saw what Insilico could do as a provider, right, so we started synthesizing our chemistry at his company and he saw that and it was pretty magical at that time. And he decided to join the company.  

Alex Zhavoronkov: And said that, look, I don’t want to provide services. I want to discover and, you know, take advantage of this technology, which is pretty much like as a consumer today, if you see how ChatGPT performs, right, some certain tasks, you also want to get into the field. And back then we thought about, okay, what is a central theme of Insilico? Just like what you’ve just asked in the context of very rapidly changing technology, because technology is changing much faster than we think than we see. So by the time we finished talking, somebody is going to publish a generative AI paper somewhere that will have some impact. And at the same time, drug development and discovery is very slow, right? So once you start the experiment, very often you need to wait for a very long time until it completes. So the pace of innovation in AI is, of course, much faster than the pace of innovation in drug development and discovery. So how do we cope with that? And we thought, well, you know, let’s forget about technology and drug development and discovery and let’s think about who we are and why are we here for. And let’s come up with a set of values that will be with the company for as long as it exists. And we want to be in this game for a very long time. As you know, we’re also focusing on longevity and those values where we actually brought a team together. So we had an offsite event with all employees and we decided that we exist for three purposes and we have three fundamental values.  

What is drug development and discovery?

Drug development and discovery refers to the process of creating new medications or improving upon existing ones. This process typically involves a number of stages, from identifying potential targets for drug action, to conducting preclinical testing in animals, to conducting clinical trials in humans, and ultimately seeking regulatory approval from agencies like the FDA.

The process of drug development and discovery often starts with identifying a target molecule or biological pathway that is involved in a disease or condition. Researchers may then use a variety of methods to identify compounds that can interact with this target, such as high-throughput screening of large libraries of molecules or computer-based modeling and simulation.

Once a potential drug candidate has been identified, it will undergo preclinical testing to evaluate its safety and efficacy in animals. If the results of these studies are promising, the drug will then move on to clinical trials in humans, which typically involve several phases of testing to evaluate safety and efficacy in increasingly larger groups of people.

If a drug is found to be safe and effective in clinical trials, it can then be submitted for regulatory approval, which involves demonstrating that the drug meets certain standards for safety and efficacy. Once approved, the drug can be made available for use by patients, subject to appropriate prescribing and monitoring practices.

Alex Zhavoronkov: One is patient first. So whatever technology is out there, whatever the investor interests are out there, whatever stakeholders are out there, the most important stakeholder is the patient. And we need to figure out a way to deliver the drug to the patient as quickly as possible, regardless of the means. So second is relentless innovation. So we want to ensure that we do not stop innovating regardless of the pace of change in technology. And right now we saw several cases where our software started performing so consistently well that people get addicted to it and we don’t want to change the engines. Right. And that’s kind of what happened to Google recently, right? You get addicted to Google, you still will search Google. But with the advent of generative tools, you now want to change. Right? And in our case, we really want to see those generative tools and many other tools integrated into the platform so you cannot stop. So what? Our fundamental value is relentless innovation. We will not stop. That’s why we have two CEOs in the company. Right? Just to facilitate for this, for this value. And the third value is transparency and integrity. So we want to ensure that everything that we claim is backed up by facts and we try to have a few research papers usually. So academic papers supporting the claim. So two or three peer reviewers from academia or industry would be able to back this claim because it’s peer reviewed. And those are the three fundamental values. And they basically define what the company is and what we do. 

Alex Zhavoronkov: And that’s why we have two CEOs. So Dr. Ren, after he joined, he joined as a chief science officer, but then very rapidly demonstrated that he can outpace pretty much any other biotech company. Right? So we nominated two pre-clinical candidates within eight months after he joined the company. So that’s pretty cool. And last year we nominated nine pre-clinical candidates. So 2022 just for reference, big pharmaceutical companies, if you look at just small molecule drug development and discovery and just internal fully internal R&D, they usually nominate like four to five, to maybe sometimes ten, if it’s a very best year in the world in history. But usually it’s less. And it basically means that on a very tiny fraction of the budget we managed to perform at the level of a big pharma company and we tried to go after more novel targets and Dr. Ren became a co-CEO. So we decided that, okay, we are going to have two heads. One is going to be AI to facilitate for relentless innovation in AI and also robotics. And the second head is discovery and also contribution to AI from the drug development and discovery perspective. So that is, to answer your question, so how did we develop in software and development and discovery at the same time? So we have two CEOs. I’m in charge of tech and Dr. Ren is in charge of the classical kind of backbone of Insilico, where the value is created by going after new targets or commercially tractable targets and doing your own drug development and discovery. 

Harry Glorikian: So to jump on that or to build on that. Right. You guys got some exciting news in February when the FDA granted Insilico orphan drug designation for a drug you’re testing as a treatment for idiopathic pulmonary fibrosis. And I believe it’s called INS 018 underscore 055. I love all the the code names we we give our drugs. Right. And my understanding is that you discovered it using your own AI for finding small molecule drugs. And it’s one of the ones that’s farthest along in your pipeline. So could you tell me the short story behind that drug as a way of illustrating the company’s work overall? 

Alex Zhavoronkov: So sure. It’s kind of sad that I cannot use the visual aid to share a screen. But basically in around 2016, we published our first research paper where we demonstrate the use of generative adversarial networks for drug development and discovery. We then for several years perfected that system. And in parallel, we also worked on generative biology and all kinds of deep neural network approaches to identification of novel targets. And in 2019, we started releasing the software to the market in and then we figured out that, look, now if we want to go after something big and also to promote our software, we need to make a big demo of how the software would perform in real world context. And we also noticed that many users of the software were very skeptical. They usually do not want to bet on new, newer targets or synthesize more interesting molecules, and they needed to have a real-life demo. So we raised some significant amount of funding, originally it was $37 million, and decided to go after our own project and our own drug development and discovery. So we decided to go after, my passion is aging, so I wanted to find something that has dual purpose, so age related, but at the same time commercially tractable and purchasable in a specific disease. So I decided to go after fibrosis specifically as a biological process. It’s not a disease by itself, right? You need fibrosis when you are younger or you need to heal some wounds or just, you know, have a patch in a tissue. 

Alex Zhavoronkov: But when is too much of it? When there are too many patches, systems break. Right? And you want to ensure that there is no excessive fibrosis. So we decided to go to, we linked it to aging. So we built massive models of aging back in 2016-17 using gene expression data and other data types found a few targets that are implicated in aging and fibrosis. And then we create quite a bit of context, biological context behind those targets we could test. We prioritized 20, we could test only five. All five worked and we prioritized one. And then we decided, okay, well we now have a demo to show. Well, do we want to release it right away? And at that time we already started publishing a bunch of other things for target discovery, many proof of concept. And we noticed that people are not necessarily, you know, buying into this early stage research and using the software more or trusting the software more. We needed to go further. So we thought, well, why don’t we also use our generative chemistry AI, which can generate small molecules with the desired properties for that specific target. And why don’t we continue this into the real therapeutic program? So we synthesized a few molecules, tested, got really nice nanomolar inhibitors and then we optimized several times, got those molecules to be very advanced. So basically two pre-clinical candidate and we did it on only a few rounds of synthesis. 

Alex Zhavoronkov: We synthesized just 79 molecules for that program and most of the molecules were hits and many were amazing hits, right? So single digit nanomolar and selective and with many other properties, orally accessible, metabolically stable, basically up to 32 properties were met. And we did a few efficacy experiments. So in liver fibrosis, in kidney fibrosis, in lung fibrosis and skin fibrosis. And in lung fibrosis, we’ve managed to achieve a pre-clinical candidate for idiopathic pulmonary fibrosis. So we did multiple experiments on human cells, patient derived cell culture, then several mice models. Mouse models are usually, those were bleomycin induced mouse models. So you induce fibrosis and then show how you rescue it with your own drug. And we did this very, very quickly. In parallel, we have achieved a preclinical candidate for kidney fibrosis and also almost achieved a pre-clinical candidate for skin fibrosis. And when we thought, okay, well, will this be enough for people to believe that AI is outperforming humans. Not outperforming humans, but outperforming the traditional methods. And we did a quick survey and realized that, okay, well, most of the companies would like to see some clinical data. So we decided to go all the way and did IND enabling studies, GLP tox, and the entire package, we did that reasonably quickly. But there you cannot move quicker than the than the contract research organization can provide you with the data. You have to follow all the protocols, right, because the FDA will require it. 

Alex Zhavoronkov: And we actually did a little bit more because we wanted to have several redundant data packages. And since the target was new, we decided to also do a small pilot phase zero study in humans. That was November 2021, and we completed a very small eight healthy volunteer study in Australia, looked at how the compound is distributed, did very, very basic safety, and then we decided to go phase one, full phase one. So 80 healthy volunteers and started at an announced that in February 2022. And just recently we’ve announced the completion and receipt that we received the top line data from this phase one. So we are hopefully going to start phase two soon in the US and in China. And then we applied for the orphan drug designation and got that from the FDA. So we’re very, very happy about that because it provides a lot of benefits for the program, including it becomes more investable, right? Because this Inflation Reduction Act has many negative effects on small molecule drugs, right? So companies do not really have a lot of luxury to invest in innovation. And we are probably the most innovative program, in my opinion, out there, period. Right. And the fact that we got is great, right? Because now we can say that there are additional benefits associated with this program and maybe we can have longer market exclusivity after approval. So that’s that’s the major benefit from the ODD [orphan drug designation]. 

Harry Glorikian: So, you know, you almost have to thank those guys that didn’t adopt the technology earlier to force you sort of to go down this path. Right? It’s the Clay Christiansen Innovator’s Dilemma, which is they’re risk averse and they say we’ve got enough money that we’re willing to wait. But that pushes a few companies along to really, you know, and at some point you may not even need them, right, because of the engine that you’ve created. 

Alex Zhavoronkov: We need them! We need big pharma very, very much. Somebody needs to sell the drugs and somebody needs to do the review of the data. So yeah. 

Harry Glorikian: So I want to go back to 2014, right? You know, the company got started. What did you, what did the market need? What opening did you see? I mean, what were the unique technologies that you had developed that you thought could address those needs? And I mean, those technologies have advanced orders of magnitude since 2014. I remember looking at it back when we were first talking about it compared to where we are now. And they’re not even, I don’t even know how you draw the curve for how much it’s changed. But walk me through the beginning days of like, what did you see? What did you start with? And. How are you constantly innovating to keep up with this? Because this is the big question for every company. 

Alex Zhavoronkov: So sure. My background is in computing, right? So computer science. And then I worked for a company for a number of years and that gave me a little bit of a competitive edge when the deep learning revolution started in 2014. So back then I was already in biotech, so I switched into biotech from GPU computing around 2004. So 18 years ago, Well, 19 years ago. And computing is necessary for deep neural networks. Right? You have GPUs, general purpose graphics processing units that allow you to perform many highly parallel computational intensive tasks much, much better than on CPUs. So most of the deep neural networks in existence today, they are trained either on GPUs or on specific semiconductors that are made for deep neural networks. Now you have those as well. But previously it was GPUs. And after I quit the semiconductor business and decided to focus on biotech and did my grad work at MSU, that my grad work at Johns Hopkins, worked in industry for a while and a lot of my friends from ATI Technologies migrated into Nvidia. And Nvidia, as you know, is the technology giant which is the leader in anything, anything AI. So around 2013, one of my friends called me up and said, Well, you know what, Alex, you should come in and give a talk at Nvidia because you know you are in deep learning, right? And you are in GPU where a company we want to expand into life sciences. 

Alex Zhavoronkov: So I think that at that time. So Jensen Huang, the CEO of Nvidia, is probably the most visionary technologist, entrepreneur and just beautiful human being who built this industry. He already back then thought about the applications of deep learning to healthcare, and they wanted to learn from the experts. So I came in, I gave a talk. 2014 and 2015, we started participating in one of those Nvidia GTC competitions, became a finalist. At that time, most people didn’t get what we were doing because we were all also trying to, you know, whatever I do is focused on aging because I think that’s the ultimate killer and ultimate biological process that leads to many diseases. And if you don’t understand aging, you probably wouldn’t be able to understand many diseases as well, right? So you’re just going to be looking at individual targets and some very primitive mechanisms without, you know, understanding the global picture. And our idea back then was to train deep neural networks, massive, deep neural networks to predict human biological age using any data type and using age as the as the feature that we predict, right, because we change in time. 

Alex Zhavoronkov: So the idea was to look at biology in time. So it’s actually pretty credible, but at that time we didn’t know how to pitch it properly. So we struggled for a couple of years to fundraise because back then nobody would want to invest in an idea like that. Now the industry actually kind of emerged from that, the longevity industry, but back then you would not be able to fundraise for something like that. So I had to put everything I had myself. So I even sold my apartment and put the money into Insilico to demonstrate that deep neural networks can not only outperform traditional approaches, but it’s a completely different way to do things, period. Right? So we actually back then I started hiring people and many of them demonstrated that, you know, gradient boosting learning is outperforming deep neural networks in certain tasks or support vector machines are great for certain tasks. But I decided back then that, look, we’re actually going to use deep neural networks, period, right? And just bench when we use something internally has to be deep neural net based. It’s not, it doesn’t make sense. But trust me, in the future it will. Right? And turns out it was a very good bet because once generative adversarial networks were published by Ian Goodfellow and that was 2014 together with Yoshua Bengio, that was the cusp of the deep learning revolution, right? So first deep neural basic deep neural networks, convolutional neural networks and feedforward neural networks outperforming humans in image recognition, voice recognition, text recognition, many other tasks and generative models got published. Once they got published, I decided to immediately jump onto the new technology, and we wanted to use GANs [generative adversarial networks] everywhere, right? So if you look at the literature, media, whatever, we met back then, I was already preaching GAN. Right. 

Harry Glorikian: Yep. 

Alex Zhavoronkov: And we thought that, look, we are going to use GANs and to try to generate synthetic biological data, small molecules with the desired properties. Try to forecast into the future some clinical trial designs and clinical trials, and also generate entire human populations using age as a generation condition. And back then, a lot of people just didn’t get it. So it published some really cool papers, filed some really cool patents. So that’s actually another story I can talk about. But, uh, investors, serious investors didn’t get it right, and only in 2017, when I first synthesized the molecules and tested them at Wuxi Apptec, Wuxi Apptec saw the results. So they saw real chemistry and they said that, look, it’s actually pretty primitive now and some of the molecules are not synthesizable or good. But after testing, they got convinced. They invested and led our Series A in 2018. So that’s when we actually started getting capital attention. And with Wuxi’s support, once they joined the board, they basically introduced a lot of quote unquote Big Boy features into the company, right? So it became a much credible, much more credible player in drug development and discovery. Right? Because they there is a, Wuxi also carries a lot of weight. And they have experienced drug hunters, medicinal chemists, huge number of synthetic chemists, biologists. So we piggybacked on that company quite a bit. And later, once we decided to introduce the software, we of course, got a bunch of other investors who wanted to make it happen and also deliver this software to the community. So it’s good for everybody. And you know, coming back to 2014, we never imagined where we are going to be today. But the bet on relentless innovation and constant change was a very good bet. That’s why, again, we have two CEOs today. So I can continue doing what I was doing and being quote unquote, you know, the technology cowboy. This attitude is very often, uh, you know, people don’t like this kind of attitude in pharma. I know. Yeah. Especially when…. 

Harry Glorikian: I know, I have people around me. Yeah, I know. I have people around me like this all the time, and I just, I don’t because the turns of technology are becoming so compressed that if you’re not looking and not moving, you it. I don’t think people fully appreciate how quickly things are changing. And sometimes I can’t read enough or fast enough to keep up with what’s happening to sort of be able to draw a line of where the puck is actually going. 

Alex Zhavoronkov: Exactly. And at that time, we also faced a lot of resistance. So, you know, some of the people who are deeply admired, I still respect them a lot. But back then I admired these people. They were like heroes of mine, right? They built a drug development and discovery companies based on computational approaches of the past. We actually still use those approaches even today. And they work, they just go slower. And you can combine them as you know in your big Lego constructor you can still fit them in, but they are not the main driver of our engine. So many of those former heroes of mine, they actually started criticizing what we do, right? So because in drug development and discovery, very rarely you get the drug on the market. Right? And very rarely you get the drug in the clinic from computational approaches. Right. So that makes a lot of people very skeptical. And there are only a few key opinion leaders who actually have done that. So have taken a drug from the computational stage all the way into, you know, phase two clinical trials, uh, even for non targets or something that is novel. There are, you know, I know only a couple and back then many of them are started criticizing generative approaches, right? And saying, look, it’s an incremental change. You know, it’s not going to be a paradigm shift anywhere. Uh, show us the drugs in phase two or past phase two, you are going after two easy targets, the molecules are not novel. Oh, you cannot replicate that easily because the model is trained. 

Alex Zhavoronkov: And we don’t know what’s in the dataset. And you need to dumb it down to like one primitive model instead of a huge Lego system to demonstrate, you know, that that small piece of Lego can actually do magic. No, it can’t. And we struggled quite a bit. And once those people started openly criticizing and even writing papers on that, I got really frustrated. Right. And now, of course, you know, I still respect them, but they are not my heroes anymore. You know what I think is the best criticism—it’s a demo of superior performance with your own model, right? So we try to compete on the outputs of technology and technology itself and demonstrating that you can develop better tech rather than trying to criticize somebody who came up with a better tech, even if the tech is still uncertain. Right? And you see, so one of my role models is, for example, Elon Musk, right? So you cannot take it away from him. Five companies, two of them are delivering at scale, right? So you are driving his cars, you are using his satellites for communication, right. And payload delivery. So but very many people still criticize him saying, you know, it’s a farce, fraud, bluff. And that will continue, especially people who are in the industry for a very long time. And they might say that, you know, he piggybacked on the government to get where he is today. You know, innovation is incremental. You know, you got the statements all the time and you won’t be able to get around it. Those people just need to… Well, that’s where aging actually is beneficial for a society. Um, so yeah, so they won’t get our treatments if they live to see them. 

Alex Zhavoronkov: But yeah, I was very frustrated and we had to demonstrate clinical stage assets and also go after very difficult targets to prove them wrong in the investors eyes. Right. Because your main source of existence in this industry are the investors, right? And if your investors are people who have been in this industry for a very long time and they of course see this criticism from those people who are heroes of mine. Right. And they are very well respected. Very often they also have doubts and questions and they have to come up with their own teams to evaluate the performance of your drugs and the performance of your technology and software. And you actually need to show how the system delivers. They need to look at the molecules, they need to look at the algorithms. But it’s a little different than, you know, publishing a research paper because they can use the software, right, that you provide, right? Um, and it made our, our lives more difficult, this criticism. But after, you know, we demonstrated the experimental validation on more difficult targets and, um, significant speed ups in more common applications. And then our partners generated, synthesized, tested and got into late stages. Uh, investors actually, you know, like to tap each other on the shoulder and say, Yeah, well, we knew it was, was great. Right? So now generative AI has been consumerized, right? And of course a lot of people are criticizing the systems. Even myself, I basically published an opinion paper recently saying that, look, just generative AI, we love it, but don’t use it for self-medication, for example. Right? So people just need to sort of ,physicians need to advise their patients that, look, some of those generative systems are outstanding. You are relying on it. However, don’t use it for self-medication today. Actually, talk to my head of communications. And she’s saying, Alex, you know what? We went to ChatGPT and asked which company was the first company to use generative AI for drug development and discovery? And it shows us to be the first one and the main one and says all those positive things. And I’m like, do you realize what you are doing? You are now utilizing ChatGPT as your primary source of information and you’re very happy about that. I mean. I’m very happy about that. But you know, if you are quoting that as the ultimate arbiter, I think that you should be probably doing a little bit more research into why it got into this. Uh, you gave you this argument and I’m of course very happy that it does. But I think that if people rely on those generative systems at the consumer level too much of this output needs to be benchmarked, validated, and we need to ensure that at the patient level, we do not cause any harm. 

[musical interlude] 

Harry Glorikian: Let’s pause the conversation for a minute to talk about one small but important thing you can do, to help keep the podcast going. And that’s leave a rating and a review for the show on Apple Podcasts. 

All you have to do is open the Apple Podcasts app on your smartphone, search for The Harry Glorikian Show, and scroll down to the Ratings & Reviews section. Tap the stars to rate the show, and then tap the link that says Write a Review to leave your comments.  

It’ll only take a minute, but you’ll be doing a lot to help other listeners discover the show. 

And one more thing. If you like the interviews we do here on the show I know you’ll like my new book, The Future You: How Artificial Intelligence Can Help You Get Healthier, Stress Less, and Live Longer.  

It’s a friendly and accessible tour of all the ways today’s information technologies are helping us diagnose diseases faster, treat them more precisely, and create personalized diet and exercise programs to prevent them in the first place. 

The book is now available in print and ebook formats. Just go to Amazon or Barnes & Noble and search for The Future You by Harry Glorikian. 

And now, back to the show. 

[musical interlude] 

Harry Glorikian: So, Alex, I want to get into the two tools you guys use. Your pharma AI platform consists of two different parts. One is, I believe, called PandaOmics, and the other one is called Chemistry42. I’m wondering if you can walk us through, you know, what each of those are designed to do. How do they complement each other? I mean, maybe you can start with PandaOmics. And then I want to dive deeper into the Chemistry42 specifically because it uses a generative AI approach. And as you’ve mentioned, like lately, you know, you know, GANs have been, oh my God, the rage, right? In tools like Dall-E and Stable Diffusion in art and ChatGPT in natural languages, right? I want to understand the two systems and then what are the similarities and differences are between Chemistry42 and those other forms of generative AI that we’ve been hearing about. 

Alex Zhavoronkov:  Sure. There are actually three systems that we have. So there is PandaOmics, Chemistry42 and Inclinico. And all of them are using generative AI in one form or another. Chemistry42 is most generative, so to speak.So you can trace the papers back to 2016. So I’ll walk you through those three systems. So PandaOmics, it’s a system which allows you to identify novel targets for a broad range of diseases. We perform best in tissue-specific diseases where large amounts of ’omics data is available. So what this system does, it utilizes different data types. So either biological data types like gene expression, whole genome sequencing and methylation. So those are the three predominant data types that the system works with to come up with novel target hypotheses based on more than 60 target discovery philosophies. 

Alex Zhavoronkov: So you can come up with a target if it’s genomically driven, right? So if there is genetic evidence that the target that the gene is implicated in disease and it’s somehow causal, this system has a philosophy for that. Um, if you rely more on expression, so relative levels of expression of a certain gene, change between normal and disease or between different types of disease, um, the system will be able to capture that, and also use it as a way to rank the target and implicate it in the disease. And there are many, many others, right? So again, more than 60 philosophies and you only acquire those philosophies over time by working with key opinion leaders that discover targets or big pharmaceutical companies or biotechnology companies. Or venture capitalists, actually, as well. So many of them, they have different ways to explore the target space. And this system, in addition to ’omics, it also has massive amounts of text data. So we estimate that the system sits on around $1.5 to $2 trillion worth of data because the way we capture this data, we track biomedical grants and you can trace my publications on that to 2011. So by the time the grant is published or given to the scientist, you’ve got an abstract for the grant and some basic information about the idea, the concept that the scientist is working on. We usually link this grant to publications later on, to patents and onto clinical trials and unmarketed products and even some real world evidence. So that’s what we do in a very careful, curated way. And we try to trace the ’omics repositories that stem from this grant as well in publicly available databases. And this ability to look at the propagation of knowledge creation and data creation from the time you get the grant all the way to the market product is the core of what Insilico is in general. 

Alex Zhavoronkov: So we look at everything thing in time. So we acquire data in time as well. And we look at the biological data in the system as the primary source of novelty for targets, right? Because many targets are out there, they’re very well known, but they are very risky. Right? So the novel, the more novel the target, the higher the risk. And then we look in the literature patents and many, many other data sources for more pockets of evidence. So there are different AIs, including generative AIs, that start building knowledge, graphs and knowledge, the target dependencies and different causal models on trying to implicate this target in a disease further. And of course, if the target has been clinically proven to work, that’s very validated, but that’s it also is demonetized, right. So you should not develop chemical matter for something that is extremely well understood and there are lots of good molecules against it, right? Investors just won’t give you the money for that. So upon the ’omics allows you to make those decisions on which target is implicated in disease, what is driving the disease and also even some commercial tractability whether you should proceed with this target or not. If you want to build a drug development and discovery pipeline for a specific therapeutic area. So PandaOmics solves that, and it’s so easy to use that the high school students can use it. And they do. So we have cases where high school students from Norway and the US, used PandaOmics to identify targets and diseases like glioblastoma and age associated conditions and presented posters. And now we know that one of the groups submitted a paper for the first time with authors going, you know, as early as 16. So imagine, back then what were you doing when you were 16, right? I remember. I didn’t know basic biology. 

Harry Glorikian: That’s awesome. I wasn’t doing this. No. 

Alex Zhavoronkov: And so it gives students superpowers. And then Chemistry42 allows you to prosecute those targets that you identify using PandaOmics with novel chemistry. And that’s a very cool generative system which can take either a ligand, so an already known molecule and then redesign it with the desired characteristics. So pretty much if you were to think generative imaging, imagine if you got a picture of Brad Pitt, but you wanted to see Brad Pitt when he is 80, female, and with glasses and with some Asian characteristics. And a generative system would be able to produce a wide range of Brad Pitts with those features. Right. But still retaining some features from the original template. So we can do the same with chemistry, then that’s a very rare application of our tool, mostly used for lead optimization. So when you already have a starting point and you want to make it better, the tool is for you that that pipeline in the tool is for you. 

Alex Zhavoronkov: The most common application is when you start with a target. So you give it a crystal structure. So you crystallize the protein, you identify a pocket within this protein where you want the molecule to bind. Or the system does it automatically for you. Then you annotate this pocket or the system automatically does it for you in the autopilot mode. And then within this pocket, the system generates molecules with the desired properties. And those properties you select, pretty much like you would say in natural language to the system that want to see a drug-like molecule that inhibits this protein target, binds to this active site or some other site on the molecule in a protein with this affinity. And at the same time, it must bind only to this target and not anything else or find something else very weakly. It needs to be penetrable to a specific tissue, for example, cross blood brain barrier and it needs to be metabolically stable. And we want to ensure that you can take it orally, right? Not inject it. So it needs to be stable enough for your gut not to digest it too quickly or your liver to digest it too quickly. And many, many, many other parameters. So there are many, many parameters. A huge number of parameters that you can select for your molecule. And then the system starts generating using many, many generative engines. There are more than 40 generators in there using different approaches. Either it’s GANs or it is transformer neural networks or it is genetic algorithms and they all start generating molecules with the desired conditions. 

Alex Zhavoronkov: And then we have a massive predictive system. So where you have more than 500 different ways to evaluate the molecules that are being generated by the generative system. And evaluate them, whether they are good or not. So using different approaches in 2D, in 3D, if you have a little bit more time and money, you can also evaluate them using other representations of molecular structure. Even go back to go down to the physics layer. And this predictive system is either rewarding or punishing each individual model. Generative model. And the ensemble. So you’ve got the reinforcement learning system going on, right? So if the molecules that are generated by the system with the desired properties are bad, so they might be toxic, they might be not synthesizable, they may have many, many other properties that you don’t want to have, the model gets punished and it learns over time to generate really, really good molecules for that specific target with the desired properties. 

Alex Zhavoronkov: And at the end you also can use this predictive system to rank the molecules that at the end are generated by this ensemble. And say, okay, I want to synthesize the top ten or no, I want more diversity. I want to synthesize, you know, some molecules with this and this and this scaffold. And you can filter out and then  send them for a synthesis. So in the first pass, you usually generate and synthesize about 20 molecules, and many of them would be hits depending on the target type and depending on how lucky you are or what your generation conditions or filtering conditions are. So you can get way outside the known chemical space with Chemistry42. You don’t need to screen the libraries to identify the molecules that you want. You can generate a lot of those great molecules from scratch. So instead of searching for a needle in the haystack, you can create a haystack of great needles. And then prioritize them for your tasks. 

Alex Zhavoronkov: And then we have a system called Inclinico. This system actually incorporates some engines from PandaOmics and it’s predominantly used by and we are piloting currently with many hedge funds and banks and capitalists who actually bet on small and medium type biotechs that have clinical stage assets going through phase one to phase two, phase two to phase three. So we are not that accurate predicting phase three outcomes. But phase two outcomes, when the trial is for a small molecule targeted first in class single agent therapy, we do really well. So about 20% of the trials that we look at are predictable with our algorithms with reasonable accuracy. And I’m talking about, you know, 87 to 92%, very high accuracy. And we use two scores to evaluate the probability of transition from phase two to phase three. So one is target choice score. So we try to implicate the target in a disease and then look at how well did the biotech company conducting the trial select the target? And then we look at how heterogeneous the target is in a patient subpopulation. So we utilize a few modules from PandaOmics for that and also some modules, that were trained specifically on clinical trial data. And then so that’s one score and then another score is clinical study design. So there we utilize a generative approach. We use a transformer neural network that can extrapolate the trial into the future and also tell you the probability of success of that specific trial based on clinical study design. And that system is also interpretable. So we can see which features are more likely to contribute to the probability of success or failure of this specific trial. 

Alex Zhavoronkov: So and we turned the system originally it was actually it was one of the most, one of the oldest projects in Insilico, six and a half to seven years old. The first paper in this field we put on archive or just on expression data in 2016. But then we did, we approached the problem from many angles, looked at many different approaches on how to score trials and performed quasi prospective validations. When you train… up until 2015 to 2020, for example, when the system did not see the trials or even traces of trials because it’s trained on everything before those trials even started, and then we realized that that is not enough. We need to train on, we need to produce prospective validation. So we train on everything. Up until today, we make predictions and then we wait for one or two or three years even, or five years and wait until our predictions are the clinical trials were predicted. Read out and see if we did it well. And now we are getting a lot of this prospective validation. So we started that again back in 2016. We can find some of our papers on archives on preprint servers where we can date stamp, right? And there is a prediction and we cannot remove them from there to see how well we did. And in most cases we did extremely well. So yeah, after we validated to our for ourselves and also we have virtual portfolios on Yahoo finance and other resources showing that we can even outperform biotechnology analysts using this tool in terms of, for example, if you just invest $1 million into an equally weighted portfolio of trials that are scored using our algorithms, it gives you, I’m not going to make claims here, but it gives you decent returns. 

Harry Glorikian: No, I’ve been looking into this. I mean, it’s, you know, A, you can have a better probability of prediction, but B, you can actually tell the person doing the trial how to make changes to improve the trial, to improve the probability of success, which they may not consider, because the machine is sort of looking at many different parameters to help you improve the outcome. 

Alex Zhavoronkov: Exactly. And in my case, what I do also now, I try to bridge Inclinico and PandaOmics so those two systems can actually be combined for better target discovery and target validation. And now those two systems run my fully automated robotics lab. So I’ve got a fully automated robotics lab in Suzhou. It’s an entire floor full of robots that are interconnected and humans do not go into the lab. We just throw in a sample, it gets processed. And if you like, I can tell you. About how it does it. 

Harry Glorikian: Yeah. We’re gonna get to that in a second. 

Alex Zhavoronkov: Yeah. The data is fed into one. PandaOmics and Inclinico Those systems select the target and then the system does what it does. 

Harry Glorikian: So. Let’s step out of this for one second. There’s been a huge uptick of new companies founded, the amount of venture capital raised around AI for drug development and discovery. I mean, it’s starting to become very blurry. I mean, you found it Insilico in 2014 before that really big wave got rolling. So arguably you were ahead of the game, right? But in in past conversations with the media, I think you’ve predicted there’s going to be a shakeout, a consolidation in the area that, you know, a lot of these companies may go away. I think in 2021 you told Max Gellman at Endpoint News that there was going to be a shakeout in the next, say, 12 to 24 months. And if you run the clock forward to today, that 12 to 24 months puts us right, right about now. Right. So let me ask you a few questions. Like, do you stand by that same prediction? You know, if you if you had to, what would you change about it? And if AI in pharma is in a kind of investing bubble, you know, what do you see as the fundamental weaknesses that are going to get exposed or the assumptions that are going to turn out to be wrong? For example, are there just too many me-too or copycat companies that don’t have valuable original ideas about how to use AI in drug development and discovery? Let’s start there. 

Alex Zhavoronkov: So sure. So first of all, I think that my prediction came true. And much earlier than we expected. Because in 2022, that was the real shake up year. So several of our, not competitors, actually don’t treat anybody as a competitor unless you are in the same target space, right, because if you have your own discovery programs, usually you are if you are on the same target space, then you are a competitor. But if you are not on the same target space, you are just competing for attention. And we never had too little of it, right? So, many companies have collapsed. They went out of business. And currently the market has shrunk dramatically. And the trend continues. So you saw so Endpoints just recently published an article that well, one of the companies in California, I’m not going to say which one because actually like the guys a lot, they just downsized, cut 30 plus percent of their R&D of their staff, right. And all scientists. So now they focus on drug development and discovery, trying to get some of their drugs that they already discovered using AI into the clinic. And that’s a tough choice, right? So you basically move away from being an AI company to becoming a biotech company. That’s the choice that we avoided by having two CEOs. Uh, you know back way then, uh, when we realized that, you know, technology moves faster than biology, than biotechnology. And many companies went out of business completely. So some went bankrupt. Most downsized. And, um, I have not seen many being created. If 2019, 2020, that was the peak when huge numbers of companies were created, many of them are not around anymore. 

Alex Zhavoronkov: And there is a good reason for that. So first of all, if you are to partner with a pharma company today, you need to have a very high standard for what you are showing to them, because most of them, they have internal groups that have ramped up. That know that what they are doing right, they will look at your algorithm and they will either copy this. algorithm and say, Thank you very much. We didn’t know that it works, right? And that happened to us many times when we were kind of smaller and less experienced. Or they would give you a pilot. And that pilot will need to result in certain, you know, synthesizable and active molecules, preferably that could be progressed into a working program, right? Therapeutic programs. So most of the pharma companies are now smart enough not to partner with companies that will not be able to deliver and that will not be around, you know, two years from now. So it’s very stupid for any pharma companies to partner with, you know, Series A-stage companies, right? Because the probability that they will be able to raise Series B and continue supporting big pharma is very low. And became lower. Because we’re also in a biotech winter. So many investors in AI are shying away from that. 

Alex Zhavoronkov: And that’s the second argument. So one of the main reasons why many AI companies are dying is because investors either ask for much lower valuations to put the money in because they understand, okay, worst case, they will be a software vendor. And the software market is tiny, right? So actually the platform market is not huge. I would say it’s you know, if it’s $1 billion globally, I would be very surprised. I’m talking about like pure AI powered drug development and discovery software. And many investors would give you very low valuation. Many of them they will bring very experienced analysts. And they will scrutinize the technology and it will be much more difficult to fundraise because most of the time we have already had a footprint in that technology because we explored it very broadly. And several other companies well. And some are exploring those right now. So that’s the second. 

Alex Zhavoronkov: Third is both investors and pharma companies and even employees, potential employees, like to have see clinical stage assets from your AI engine. If you don’t have some, that tangibly came out of your engine and are in, you know, phase one, phase two, they don’t believe in the company very much. Right. Because. again, they understand it’s the biotech winter. Job security or the company’s future is very murky when the markets are not flying high and the investors in pharma are much more educated. So right now we see a lot of collapses. And I think that this trend is going to continue. Like, for example, one of the brightest, brightest young scientists I met recently who, you know, did his graduate work at a very top university in a top lab. Has really amazing ideas. He was asking me for guidance, and asked, “Alex, okay, well, how do I build an AI powered drug development and discovery company today with this and this and this approach and I actually want to combine target discovery and small molecule discovery with this system and want to fundraise.” And I told him, well, you know what? For you specifically with all your background, I would recommend a completely different approach. So don’t go discovery. Don’t go too early. If you think that you can fundraise and sell this idea to investors, you should use your technology to in-license phase two or phase one clinical stage assets that are demonetized today. Right? So there are many, many companies in biotech that are collapsing that have very little cash. So go and try to work with their assets to either repurpose them or pick the best ones. And then for a specific indication, in-license, well, fundraise and in-license and take them forward, right, and take them through one stage. And that will demonstrate to the world that your algorithm works at least on the target discovery side, because you can do a better job than those biotech companies that have those trials. And you shouldn’t go too early. Right? Because A, the market for software and services is tiny. And the impact you are going to make on a field is going to be very small because you are going to be struggling and, you know, trying to crawl out of this early preclinical swamp. And most of the time you’re going to fail, even if you fundraise, even if you try to do a proof of concept. But if you demonstrate that you can do a very good job with a clinical stage asset, yes, you will require more capital, your next big thing could be the early-stage drug development and discovery, right? You can then say, okay, well, I have validated this. And let me go back to the drawing board and now show to the investors that with my own internally generated funds, after outlicensing repurposed or progressed phase to asset, phase two to phase three or phase one to phase three, and with a commercial success, I can now fund my, you know, hobby, which is early stage drug development and discovery. Because it’s gladiatorial there. It’s very difficult to crawl out. And we would need to see another hype cycle of hyper investment to see another wave of AI powered drug development and discovery companies. 

Harry Glorikian: I mean, I do see pockets where there are some unique approaches that I’ve liked. Like, you know, I was just talking to one in RNA structure and their approach. And so I think there are some there are some paths, right? But I think that some of the well-trodden paths may be more difficult, but I do believe that because of these technologies, there are some unique avenues that have also opened up that most people may not consider right off the top of their head that are possible, that are still very investable. But that’s just as an investor. I’ve got to be looking for those, you know, needles in the haystack to be able to get that next opportunity. 

Alex Zhavoronkov: Yeah, but again I don’t want to discourage anybody, right? I’m just being a realist. And I don’t want to sound like one of those people who criticized or attacked right when we were, when we just started showing some early proof of concepts for generative AI in chemistry or biology or both. And, you know, they were, of course, sounding very discouraging for no reason. Right. Just because they were jealous and this or I don’t know for what other reason. Um, or they wanted to stay relevant. I don’t want to be one of those. So, yeah, start the company. Great. You know, uh, try to make it. But my recommendation would not to do this in a traditional space, so to speak. Traditional protein targets, plus traditional small molecules, you know? Yeah. Go RNA therapeutics, go to some other therapeutic modalities, because pharma also needs new ideas in general, not only just acceleration of small molecule drug R&D. Yep. Cell therapy? Yeah. 

Harry Glorikian: So the last thing I want to touch on, you brought up a little earlier. I mean, you’re investing a lot in this sort of fully automated wet lab where robotics can handle all the experiments you need to generate sort of larger data sets and validate your drug targets. I think you said you’re building a prototype lab in Suzhou, China. This is a theme I’ve heard from many companies, right? It seems like robotic wet labs are the next big thing. What’s your goal for those facilities and what’s Insilico’s unique philosophy about how they should run? 

Alex Zhavoronkov: So sure you could, actually put those labs into two categories. One would be biology and another one is chemistry. Right. So very easy. So some companies are focused on chemistry automation and there are many. And there you should start with large contract research organizations. Like, you know, Apptec, and others. We also have automated labs, but it’s chemistry. So let’s let’s leave chemistry aside for a second. I think there’s going to be a big revolution there, but we’re not there yet. But we’re getting there. 

Alex Zhavoronkov: In biology, we’re already see a revolution, right? So and that’s where. I would say barriers of entry are going to be much higher than in automation of chemistry. So there were, so for target discovery, there are several generations of labs, laboratory robotics, So high throughput screening has been around forever. So since, you know, late 90s, companies started introducing those massive liquid handlers that pipetted, in an automated way, the molecules onto a variety of cells and cell types, cell lines, organoids, and looking at the phenotypic effect, right? So if something worked, they prosecuted that molecule. So let’s call it first generation. Second generation is that they started looking closer at the data and started generating ’omics data, including in cases where the compound did not work. So third generation, it’s more automation and more automated islands where people started integrating, for example, ’omics and imaging, but mostly put large image imaging machines together with high throughput screening machines. And companies like, you know, don’t want to single out some companies. Yeah, I’m not going to talk about companies, but some companies that basically are publicly traded today, they started back in those third generation era. And they have they focused predominantly on imaging and did some ’omics automated workflows and they have some clinical stage products right now. 

Alex Zhavoronkov: Fourth generation was when companies built predominantly by investors who started, you know, getting into this field and trying to build companies with the purpose of selling them to Big Pharma later on or taking them IPO. Um, so that’s not organic growth. I call it, you know, investor-built companies, for the purposes of accessing financial markets or selling. So their story was that we don’t have enough high quality data. We want to generate more data using robotics. And once we generated generate this data, we will find new targets. And then we will also train better AI. But currently we lack high quality data. We don’t have it. So, you know, build it and they will come the targets. And there are a few companies that do that. I don’t want to single out anybody. Uh, and, but I haven’t seen any results from those companies yet. Our approach was very different. And also in those fourth generation labs, I call it fourth generation, you still see a lot of humans walking around with microplates and, you know, shuffling those microplates into the machines. Or doing a bunch of work manually.  

Alex Zhavoronkov: So a fifth generation lab, and that’s something that I think we’ve achieved in November last year or even October last yet, is full automation. So where you create machine learning data, but there are no humans in the lab at all, right? So you basically throw in a sample and the robots do the rest. So they can do imaging all kinds of ’omics analysis and then do high throughput throughput screens and then again imaging and ’omics analysis. So we had that last year. And the sixth generation is when you actually are an AI company, a real one, organically grown, usually you don’t have clean machine learning data coming from the robot. And I actually think that it’s the wrong approach to start with. But you struggled over time and built an AI system that can discover targets without the robot. So you have enough data in the system already to be able to come up with target hypotheses yourself, right? And actually with many, many different approaches and you’re battle tested the system and now you want to take the system to the next level. You want to automate many processes to actually make your drug development and discovery faster and also remove additional bias and do it 24/7. 

Alex Zhavoronkov: So that’s what we did and that’s our sixth generation lab. And in sixth generation laboratory robotics, AI is making decisions on target choices. So it’s already pre-trained. So you have pre-trained AI, that is, that sits on top of the application layer where you have multiple models that come up with target hypotheses and based on the data it’s exposed to, it picks a target, also looks at whether the molecule is already available for this target, and selects this molecule for testing, or if there is no molecule, it can try to generate a molecule and ask us to synthesize it. Right. Or it does a CRISPR screen, right, just to see if the knockout of this particular gene will result in a desired phenotype or expected phenotype. So we have now achieved sixth generation. At least that’s what I believe in. So. Because the devil is in the detail, and unless you show a phase two, asset coming from this lab, you know all the claims are hypothetical as we know anything in drug development and discovery. So we built a lab and lets, allows you to fully automatically identify targets. Validate them without any human intervention. So you can put it in a manual or manual overdrive and select targets manually and do it also remotely. But our idea is to have a pre-trained AI that is exposed to incremental data coming from the lab. It makes decisions based on that incremental data and then it performs some action, for example, selecting molecules from libraries. Doing a high throughput screen and then doing again, many ’omics and imaging experiments to see whether the hypothesis was confirmed or not. And also it can learn a lot of biological insights from this exercise. So our lab allows you to do sample goes in, you do deep genotyping, so you understand you have imaging, you have fluorescence contrast imaging. You get transcriptomics. So RNAseq, you get methylation, you get whole genome, plus you get a few other data types that I cannot talk about. And then the system takes this data, makes decisions on the targets and on the compounds, runs a high throughput screen take, takes the compounds from the storage and then takes the so selects the cells that it wants to to run through an additional screening. And you get again high-resolution imaging. So deep phenotyping contrast imaging fluorescence. And you get transcriptomics and methylation and additional data types. After a high throughput experiment was performed. 

Harry Glorikian: And so how if you were to guess, you know. How much faster do you think the system would then be able to produce something? And then would there be a cost differential and, you know, give me some hypothetical metrics because I’m not sure if the system has produced anything yet to put it into context of how we do things manually. 

Alex Zhavoronkov: Yeah. So it produced something. And a lot. It produced a lot. Currently to run the system, it’s of course much cheaper than running it manually. Right. So if you were to go through the entire workflow, you would need a lot of scientists who are going to be working on this problems. And most likely they would have some inconsistencies. You know, I remember standing in the bench myself and pipetting into 96-well plate, even if you do basic PCR, very often, you just forget whether you, you know, put pipette at the reagent in the well. Or not. Right. Did you mix, did you not? Well, you just forget about it. Right. So especially if there is, you know, some destruction. Um, and I think that currently we’ve demonstrated that we can automate the process and cost down at least by removing humans from the lab. However, the amount of data we generate from every experiment, from every run, is infinitesimally larger than what you would expect in a normal human driven lab. And that makes it also very expensive, right? So in a high throughput screen, very rarely you do RNAseq, for example, for every sample in the plate, you would only do it selectively for, you know, samples where the compound resulted in certain morphological cellular change. That are traceable. The changes that are traceable through imaging or naked eye. And then you pick some of them, you perform those expensive ’omics experiments. In our case, we actually are hungry for data, even for samples where the compound didn’t work or it caused something, but not what you expected. So our cost per run is actually much, much more expensive currently than what you would do in a traditional manner. But if you are to look at the data economics perspective, if you were to generate the same amount of data manually, it would be, we are cheaper. But if you are just to follow the traditional approach, we are much more expensive. 

Harry Glorikian: Yeah, but the data fed into the system to then allow it to make a better decision on the next turn is significant.  

Alex Zhavoronkov: Yes, you learn. It can learn. 

Harry Glorikian: It’s going to be higher in the beginning, but technically should be more efficient and more accurate as you write. So you’ve got to invest to to get there. 

Alex Zhavoronkov: Exactly. And over time, as you get and generate more data and get better at it, and again, the devil is in experimental validation, right? And specifically in trying to see if the system can pick up novel biology that has not been known before. That is the coolest thing. And so far, that’s my goal for this year, to demonstrate that it works. And publish. Um, yeah, but the devil is always an experiment. Yeah. So our job is to demonstrate this this year. And we also want to miniaturize this lab now. Cost it down so that we can even put it again. Hypothetically, we’re not doing it right now. Theoretically, you can put it in the in the hospital. So it can do personalized medicine, something that people do anyway because they need sequencing anyway. But at the same time, the hospital would be able to do drug development and discovery. And share target lists with everybody. So that would be my dream. It’s a big long shot. So if I can give you a lecture on this right now, but, um because yeah, I actually had to live in this lab for the past, for a few months, um, to build it. Right. Because it’s not an easy exercise. And you could only build it in China. Because we rely a lot on AGVs, autonomous guided vehicles. So one of the reasons why we were in Suzhou is that’s where most of those highly intelligent AGVs are made. And many of the robotics components that we’ve integrated, they are made there, they are invented there. They are designed there. So I think that people have a very wrong idea about China nowadays. They think about China as a technology thief, but that phase has gone, right? So it’s a technology creator and now other countries are coming to China for innovation, not the other way around. And we are one of the first companies to realize this trend. And we are trying to be where this technology fountain is most abundant. Suzhou. 

Harry Glorikian: So, Alex, it’s been great having you on the show. I mean, I you know, I have a bunch of questions we didn’t even get to that I would love to maybe schedule another show to talk about, because I know you’re at the Buck Institute and talk about longevity and everything you’re doing in that space, but hopefully we can book that for a different show in the future and tantalize all the listeners with some more of your your wisdom and knowledge in the space. 

Alex Zhavoronkov: Sure. Thank you. would love to. And for everybody who is still listening, please do go to It’s a nonprofit event. The largest industry event in the biopharmaceutical industry in aging research. It’s organized by the University of Copenhagen. And it’s a five-day event at the end of August. Highly recommend it. I’m supporting it, sponsoring it, helping organizing it, organize it. It’s epic because it’s Copenhagen. And great place to to be if you want to explore Copenhagen and learn about aging research and drug development and discovery. The place to be. 

Harry Glorikian: Excellent. Thank you so much. 

Harry Glorikian: That’s it for this week’s episode.  

You can find a full transcript of this episode as well as the full archive of episodes of The Harry Glorikian Show and MoneyBall Medicine at our website.  

Just go to and click on the tab Podcasts

I’d like to thank our listeners for boosting The Harry Glorikian Show into the top two and a half percent of global podcasts. 

To make sure you’ll never miss an episode, just open Apple Podcasts or your favorite podcast player and hit follow or subscribe.  

And don’t forget to leave us a rating and review on Apple Podcasts.  

We always love to hear from listeners on Twitter, where you can find me at hglorikian. 

Thanks for listening, stay healthy, and be sure to tune in two weeks from now for our next interview. 

FAQs about Drug Development and Discovery

How does AI help drug development and discovery?

AI (Artificial Intelligence) has the potential to significantly improve drug development and discovery by accelerating the pace of research, reducing costs, and increasing the accuracy of predictions.

One of the ways AI is being used in drug development and discovery is by analyzing vast amounts of data to identify patterns and relationships that would be difficult or impossible for humans to identify on their own. For example, AI can be used to analyze large datasets of genetic information or medical records to identify potential drug targets, predict which patients are most likely to benefit from a particular treatment, or identify potential side effects of a drug.

AI is also being used to design new drugs and optimize existing ones by using machine learning algorithms to predict the properties of molecules and how they are likely to interact with target proteins. This can help researchers identify promising drug candidates more quickly and efficiently than traditional methods.

In addition, AI can be used to improve clinical trial design and recruitment by analyzing patient data to identify individuals who are most likely to benefit from a particular treatment and by predicting the likelihood of adverse events.

Overall, AI has the potential to significantly speed up the drug development and discovery process and reduce costs, ultimately leading to the discovery of new treatments and cures for a wide range of diseases and conditions.

What are the stages of drug development?

The process of drug development and discovery typically involves several stages, each of which is designed to evaluate the safety and efficacy of a potential drug candidate. The stages of drug development and discovery are as follows:

  1. Drug Development and Discovery: In this stage, potential drug candidates are identified through a variety of methods, such as high-throughput screening, computer modeling, or repurposing existing drugs. Once a potential drug candidate is identified, it is optimized to improve its potency, selectivity, and other properties.
  2. Preclinical Testing: In this stage, potential drug candidates are evaluated in animal models to assess their safety and efficacy. This stage is necessary to determine whether the drug candidate is safe for use in humans and to establish a starting dose for clinical trials.
  3. Clinical Trials: Clinical trials are conducted in humans to evaluate the safety and efficacy of a potential drug candidate. Clinical trials typically have three phases:
  • Phase 1 trials are designed to evaluate the safety of a drug candidate and to determine the maximum tolerated dose.
  • Phase 2 trials are designed to evaluate the efficacy of a drug candidate in a larger group of patients and to establish the optimal dose.
  • Phase 3 trials are designed to confirm the efficacy of a drug candidate and to establish its safety in a large, diverse population.
  1. Regulatory Review: Once a drug candidate has successfully completed clinical trials, the results are submitted to regulatory agencies, such as the FDA, for approval. The regulatory review process involves evaluating the safety and efficacy of the drug candidate and determining whether it can be approved for use in the general population.
  2. Post-Marketing Surveillance: Even after a drug is approved for use, ongoing surveillance is conducted to monitor its safety and effectiveness in the general population.

Overall, the process of drug development is a long and complex one, often taking many years and involving significant investment of time and resources. However, the rewards can be significant, as the development of new drugs can lead to improved treatments for a wide range of diseases and conditions.

What generative AI tools are working on drug development?

There are several generative AI tools that are being used in drug development and discovery. Here are a few examples:

  1. Generative Adversarial Networks (GANs): GANs are a type of AI algorithm that can generate new molecules with desired properties. This technology can be used to design new drugs and optimize existing ones.
  2. Recurrent Neural Networks (RNNs): RNNs are a type of AI algorithm that can be trained on large datasets of chemical information to predict the properties of new molecules. This technology can be used to identify potential drug candidates and optimize existing ones.
  3. Deep Learning: Deep learning is a type of AI algorithm that can be used to analyze large datasets of patient information to identify patterns and relationships that would be difficult for humans to identify. This technology can be used to predict which patients are most likely to benefit from a particular treatment and to identify potential side effects of a drug.
  4. Evolutionary Algorithms: Evolutionary algorithms are a type of AI algorithm that can simulate the process of natural selection to design new molecules with desired properties. This technology can be used to identify potential drug candidates and optimize existing ones.

Overall, these generative AI tools have the potential to significantly accelerate the drug development process by identifying new drug candidates more quickly and efficiently than traditional methods. However, further research and development are needed to fully realize the potential of these technologies.