Generative AI in Drug Discovery and Pharma, with Insilico Medicine (CXOTalk #782)

Generative AI in Drug Discovery and Pharma, with Insilico Medicine (CXOTalk #782)

Show Video

We are exploring generative AI in pharma  and drug discovery with someone who is a   true pioneer in this field who has been working  with generative AI for years. Alex Zhavoronkov   is the founder and the CEO of Insilico Medicine. What I find so interesting is generative AI is the   latest rage and hype, but you've been working in  this field for years and years, so tell us about   your work and tell us about Insilico Medicine. Our first experiments in generative AI were in   2015-2016. We started in generative chemistry,  originally, and utilized the technology  

called generative adversarial networks. That is actually at the core of the many   generative AI platforms in use today,  and that technology, to put it simply,   is a combination of two deep neural networks  competing with each other. That is why it's   called adversarial. One is generating meaningful  data in response to a query and another one is   evaluating this response to see if it's true or  false (or how close it is to the ground truth).  Since then, this technology, generative AI, has  advanced quite considerably. In 2017, Google  

published a wonderful concept where they introduce  an attention layer into generative networks. It's   called Transformer Architecture, and that changed  everything. Those attention layers allow deep   neural networks to generalize and later generate  meaningful output with the desired properties.  We've been in this field pretty much since  the very beginning. Ian Goodfellow pioneered   generative adversarial networks, so we did  not invent them. We started building on top   of them in drug discovery in 2016, and my first  paper was published in 2016 demonstrating the   applications of adversarial autoencoders to small  molecule drug design using a molecular fingerprint   representations of molecular structures. Think of it as painting or imagining new  

molecules with the desired properties just  like you do it with images today. You say,   okay, Valley or Midjourney, draw  me an image with those properties.  We thought that it would be cool to do the  same with molecules, and our first paper was   (believe it or not) 2016, so submitted in June,  published in December. It actually made some  

shockwaves in the AI community, originally,  because it was a really cool application.  In October of the same year, Alán  Aspuru-Guzik from Harvard also put   his article with a very similar idea but  with variational autoencoder on Archive,   so on a preprint server, and we were at that  time already in the peer review process in   a peer review journal. Yeah, he actually has  more citations for his paper because, later,   he published in a peer review journal in 2018. But  it went on Archive when we were still in review.  But then he joined us as an adviser anyway,  so who cares about who did it first.   [Laughter] But we were probably the first  who used generative AI to design a molecule. 

Then it took us several years, so we of course,  published multiple theoretical works around the   generation of small molecules with exact  properties. Using different approaches,   we came up with reinforced adversarial threshold  neural computer, and many, many other techniques   where we also started incorporating reinforcement  learning, so rewarding and punishing some of those   models, and published a lot of theory. In 2017, for the first time, so actually   just a year before we talked. You've raised $400 million. When   you say, "Invest the money over this time,"  you've raised a significant amount of money. 

The road to this money was not easy.  Many companies in our field, especially   those that are designed by investors to access  financial markets because it's a trend. Right?  They like to invest a lot of money right away.  Give the founder some stock (and sometimes it's   not even the founder). It's an engineered company  with very polished messaging, but usually, the   technology there is just starting to be developed. In our case, we grew organically, and we were  

founded in 2014. But only in 2019, we raised  really serious money. We raised the first $37   million (like big money) in 2019, in September. Raising this money, what is the objective?   What are you trying to accomplish? In 2017, we were poor. [Laughter] We   actually didn't have a lot of money to invest  in synthesis and just trying to explain why   generative AI in drug discovery is  different from generative AI anywhere else.  In drug discovery, you really need to synthesize  and test. The probability that you are going to   get something that works is very low because the  level of precision has to be much, much higher   than when you paint a painting that you like. The molecule has to bind to a very specific,  

very tiny site on a protein of  interest. If it doesn't, you fail.  Before that, you need to actually make this  molecule, and you make this protein. You need   to make the assay, an experimental assay. In this case, you spend maybe a few months  

building the generative model, but then you can  spend a year validating just a few molecules.   That's something for your listeners to understand. Also, the process is very expensive. It's not   just the training of the model. You need  to synthesize the molecule and test it.  Think of it as launching a spacecraft.  You have to design it. You generated   a spacecraft. Then you need to launch  it. That's the synthesis part for us.  We were cash poor at that time, so I  remember I sold my apartment at that time   and invested everything in Insilico. The molecules that we synthesized  

at WuXi AppTec (a very famous contract  research organization). That's another   thing I can probably talk about today. They synthesize tested and it worked.   In 2018, we published the first paper on that.  That led to the first kind of round of investment  

that we got from credible investors, so we got our  first $6 million after synthesizing and testing.  The first AI generated molecules after the  papers were published. In 2018, in August,   we published the first experimental  validation of AI generated molecules.  Then we, of course, advanced even more.  In 2019, we published a really big paper,  

in "Nature Biotechnology," showing that WuXi  AppTec actually decided to challenge us.  They said, "Okay, well, can we give  you any target, and how quickly can   you design small molecules and test them?" We showed that, in 46 days, we can very   rapidly come up with small molecules for the  target that they give. It wasn't a hard target.   Synthesize them and test them in many assays,  so metabolic stability, microsomal stability,   activity, so enzymatic assay, and then all the  way into mice in 46 days. That was pretty cool.  We, of course, did this experiment actually in  2018, so people knew about it and gave us the   $37 million. Nowadays, if you were to do something  like that somewhere else (in some other industry)   you can get probably $0.5 billion. [Laughter] But for us, it wasn't easy, and our first  

step after we got the first money,  round B, so $37 million in 2019,   that was actually after we talked with you.  Before that, we didn't have much money.  We developed the software that other people  can use. But not only for chemistry; also,   for biology. We allow people using  generative AI to discover new targets.   Then we also allow them and decide for the  mechanism of disease, and then generate   the small molecules with the desired properties. Then also using generative AI, using transforming   neural network, we predict the outcome of  clinical trial Phase II to Phase III, so basically   trying to replicate the entire process of  pharmaceutical R&D and turn it into a generative   exercise. We let people use the software. Then in 2019-2020, we actually thought,   "Okay, well, how do we make people  believe that it really works?   How do we make the ChatGPT moment in pharma?" We decided to actually synthesize and   test our own molecules for a novel  target, so go all the way ourselves.  

That required a very substantial capital raise,  and I can show you a slide so your readers and   listeners will understand what I'm talking about. Let me do that now, and just use a visual aid.   But I'll talk through this just in case. Here is the slide which depicts the pharmaceutical   drug discovery and development process. It comes  from a very famous research paper in 2010 by   Steven Paul in "Nature Reviews Drug Discovery,"  so I highly recommend reading this paper. It's   called "How to improve R&D productivity: the  pharmaceutical industry's grand challenge."  It shows you the many steps of drug discovery and  development starting from disease hypothesis and   target discovery, so that's where you are trying  to understand why the disease happens and what is   driving it. What are the critical components,  protein targets that are driving the disease? 

The probability of success of this exercise  or this test, it fails most of the time,   1% to 5% success rate, 95% to 99% failure  rate. Most of the time it's done in academia.   It takes decades and costs billions of  dollars, usually funded by the government.  Very often, people rely on scientific serendipity  to find a good target. We still don't understand   why Alzheimer's happens. We still don't  understand many cancers, ALS, multiple sclerosis,  

so many of those age-associated processes we  actually don't understand. That's actually one   of the reasons why I'm focusing so much  on aging because it's a big challenge.  But once you identify a target, once you validate  it also in animal models, you start chemistry   exercises. Here you've got target to hit, hit to  lead, lead optimization, pre-clinical exercises.  Here you can see, it takes you 5.5 years from the  time you found the target to the time you start   human clinical trials. Again, the failure  rate there is pretty significant, so only  

less than 50% of those are going to get there,  and it will cost you $0.5 billion by this time.  Please subscribe to our YouTube channel  and hit the subscribe button at the top   of our website so we can send you our newsletter. How is that distinct from what you're doing then?  It's not distinct. We still have  to go through all of those steps.  

We have to go through every single step. As a  matter of fact, we have to generate even more   data than usually the pharmaceutical company  would do internally because we need to learn.  Also, we try to come up with redundant  data sets. When we do an experiment,   we try to do this experiment  twice in two different labs   just to be sure and also to generate more data. If the experimental results differ from one lab   to another, we need to understand why  and also teach our AI why it happened.   It is extremely important to have this  redundancy because many of the pharmaceutical   industry failures happen because somebody did the  experiment wrong and just didn't report the data. 

We want to set the new standard in quality of the  delivery of those new molecules for new targets   so that when the pharmaceutical company looks at  it, they are like, "Okay, it's not only our level.   It's above our level (in a traditional approach." You can see that here it costs you $0.5 billion,   so AI can make this process significantly less  expensive. Also, the most important part here   is increasing the probability of success. Your major objective function in this entire   exercise (when you are looking at this slide) is  to pass Phase II human clinical trials in humans.   Phase II is when you test efficacy when you see if  the drug is not only safe but it's also effective. 

This task usually fails 66% of the time. You  can see from the slide, actually sometimes   it's an even higher probability of failure. This slide comes from 2010. Since then,   the situation actually got worse. We've got  Eroom's law working in pharma. I'm not sure   if you have ever heard about Eroom's  law, but it's Moore's law backwards. 

[Laughter] Moore's law is when things become exponential   and you've got a doubling of performance every few  years. Here you've got a reduction of performance   every few years because, in pharma, many of the  low-hanging fruits have been picked up already   and it's very difficult to find something that  is novel and, at the same time, can be done in   reasonable timeframes on a reasonable budget. We have to go expensive, and that's why many   pharmaceutical companies have to raise a lot of  capital to take the drug into the clinic. We had   to do this, too. So, we had to become a biotech. That's the reason why we raised so much money  

that you've mentioned originally. But we  raised all this money in 2021-2022, mostly,   because that's when you started  investing a lot in your own drugs.  The beauty of those drugs is that if they succeed,  if you have a Phase II complete asset and you are   addressing a chronic disease with no cure  and also potentially a blockbuster disease,   it means that after you pass Phase II,  this drug can be worth $10 billion.  We just saw one story like that, a company called  Nimbus Therapeutics. They completed a Phase II   clinical study for a very old target. I usually say that my grandmother was   working on this target. It's not  novel. The level of novelty is low. 

They completed the clinical trial for  a novel molecule for this target for   psoriasis – it's called TYK2 target –  and sold it for $6 billion to Takeda.  Yes, it takes you a long time to polish this  diamond. Yeah, very often this diamond cracks.   But if you do polish it until Phase  II, the payout is very significant.  That is how biotechnology works. When you have  an AI tool that allows you to move quickly and   increase the probability of success, the best  strategy for any AI company is to develop its   own drugs because if you can demonstrate that  you can do it, people will believe in AI. 

Pharma is very conservative. They've seen many  transformations over the years. They've seen   the human genome being sequenced. They've  seen the revolution of computing, mobile,   social networking, globalization, the emergence  of contact trace research organizations, CRISPR,   IPSC. But we only had 50 drugs  approved by the FDA last year,   and the year before was the second record year  in history in terms of the drug approvals.  In my opinion, only seven of those drugs approved  last year were innovative, small-molecule drugs.  

Sorry, but the industry is getting worse. If you have the AI tool that really is   transformative, you want to develop your own  drugs. You want to sell them to pharma. You   want to enable pharma with your own tools. But one mistake that we were making in the   past when we just started, we started  doing a lot of pilots with bit pharma.  

That's actually the topic I can expand on a lot. Why was that a mistake? That seems like a   kind of natural course of action to  partner with these larger companies.  It seems like a reasonable course of action  and, at that time, we also got lucky that   we started partnering with them in the  early days, like 2014 to 2017. We were  

always partnering on large-scale projects. At that time, the pharmaceutical companies   did not have massive AI teams themselves  (internally). They were willing to share   data. They were willing to share the experience.  They were much more collaborative. But at the  

same time, they were much more cautious  and the budgets for AI were very small.  We learned a lot during those collaborations,  so I actually decided to go end-to-end back   when we partnered with one big pharmaceutical  company. They actually challenged us in many   different departments to try to apply AI  to the most complex problems they've got.  We solved many of those problems. We realized,  "Oh, but Department A does not talk to Department   B, and Department B does not talk to Department  C. They don't even know all the processes   internally." Big pharma, it's disconnected. If we could connect it in one seamless workflow,  

we could actually increase the performance  of the pharma dramatically even by just this   connectivity, even if AI does not result in  huge performance gains. That's when the idea   of end-to-end originated. But it was a mistake  also to partner with them because what I found   as a major problem in big pharmaceutical  companies is that people come and go. 

You see this timeline. To develop  a drug and put it on the market,   it may take you a decade. But the chief  science officer of a big pharma company   or one of the top R&D people will not be  there for that long. Very often, they move.  Every time there is a new CEO, they change  R&D. You get a new chief science officer,   they change R&D. Very few projects actually mature  in this volatile, rapidly changing environment.  If you start a project, and if you don't  control it, some people change within pharma,   they go somewhere else, and the project  is either discontinued or deprioritized or   killed. We've experienced that once. I'll actually tell you. Well,  

I'm not going to tell you. Sorry. [Laughter] A big pharmaceutical company we partnered   with in 2015, the new boss comes in, kills 75%  of internal projects, starts his own agenda,   partners with his buddies, and kills many of  our projects, and we actually had in mind that   plan. We actually already partnered the plan to go  end-to-end, discover new targets, generate small   molecules, go all the way into the clinic. They had a small budget for that. There was   internal commitment. But that entire group  that did a deal with us was eliminated   within a year after the new CSO came in. Guess what. Four years later,  

he is out doing something else. Now his new  replacement is probably going to change.  We decided not to do pilots anymore. That's always the risk when a small   company partners with a large company, and then  also how can they absorb those innovations,   which is often a challenge as well. We have an interesting question from   Twitter. This is from Arsalan Khan. He wants to  know what are the negative and challenging aspects   of using AI in drug discovery. Is it inflated  expectations? What are the negative aspects? 

If you are really using AI, and if you are  developing and you are committed to that task,   there are no negative aspects.  There are only positive aspects.  The negative aspect is that you've got a lot  of very smart financers around the world,   mostly in the U.S., and actually in China  as well, who would look at the trend,   who would try to predict the trend. They say,  "Okay, well, AI is hot. I need to have an AI play.   I can come and invest in Insilico and take a small  piece of the pie and help somebody with working   technology, or I can build my own from scratch or  from starting block by somebody, so an engineer,   the company to access financial markets." What they would do, they would put a lot   of capital into the company from scratch, so  from zero, so the company doesn't need to go   through the same process we went through  where it's organically generated to try   to build from scratch. You bring a lot of  big executives, so you think, "Okay, well,  

how do I access financial markets if I don't have  the tech? I find great people, and they buy."  It becomes like football. They try to get somebody  from Google, from Stanford, from big pharma,   somebody very old from big pharma. It becomes a Tatooine. You've got many,   many different species with high profiles in one  company, in one bar. Then they start building. 

It becomes a chimera with a lot of egos where  technology takes the backseat. It's not the   main objective. Their main objective is to  get the company listed on the public market.  They try to in-license the  compounds, say, that look now we have   done it using AI. Nobody saw that AI, right?  But, yeah, we've in-licensed something.  Or very often they hype it up and say, "Oh,  I've got this new technology," or "I've got   this new idea of generating machine learnable data  using robotics. I have never discovered a target,   but I think that with this technology I can." Even before they lay the ground foundation and the  

ground floor of the lab, they go to big pharma.  Again, because the big bosses are involved,   they make a few big deals saying, "Okay,  well, here's $50 million, $25 million upfront,   and a billion dollars in arrears, and we are  going to – in five years, two years, three   years – deliver something to this big pharma." Big pharma has those budgets and, when big   bosses are involved, it's easier for them  to make those deals. The bigger the deal,   the easier it is to make (believe it or not)  because you are not trying to get a small budget   from a team that can use it internally.  You are getting it from the big company.  Very often, those deals, they later fail. The  company recognizes that "Oh, the AI company did   not deliver?" Then they think that all companies  in AI cannot deliver because, "Oh, this person was   a super-executive at Google and somebody was a big  professor at Stanford and somebody was a bigshot   in big pharma. They came together, and they  couldn't deliver. That's why AI doesn't work." 

That's the real danger of building the  company from scratch to access the financial   market to being a trend instead  of organically being in the field.  Alex, behind you is a photograph  that I believe is your lab.  [Laughter] Yeah. That's me. [Laughter] There you are. Your lab is run using   generative AI and robotics. Tell us about that. When we started that we are going to use as much   publicly available data as possible and use  generative AI to figure out how to work with   this publicly available data. This publicly  available data is usually biological data and  

also text data, is usually not very clean, and  it's not exactly designed for machine learning.   However, we've seen with ChatGPT and many  of the other generative tools that they   also take publicly available data and provide  very useful content – in imagining, in text.  It's not only about the quality of the data. It's  about the algorithm. We focused on the algorithm  

first and developed a system that can generate  small molecules on demand and also identify   novel targets on demand without the generation of  new data – just using publicly available stuff.  We've demonstrated that it works by taking the  AI-discovered target, AI-generated molecule,   all the way into human clinical trials. Now we  are starting Phase II and demonstrated it works.  Anybody who wants to refute this argument,  show me your molecule for a new target that   you've taken into the clinic. There's nobody  else that I know of that managed to do that.  I love that challenge: Show me your molecule that  you've taken into the clinic. That's awesome.  That's the new kind of way to cut through BS,  right? I think that this is the new benchmark   for companies entering this field. I think,  before you start raising a big pile of money,   you need to demonstrate that you've  at least got a pre-clinical candidate. 

Let's say you raised $5 million (if you are doing  it in the U.S. because maybe it's a little bit   more expensive). Let's say you raise $5 million  and, with $5 million, you deliver a pre-clinical   candidate. A pre-clinical candidate meaning that  you've completed at least two or three efficacy  

experiments in mice, so you cured cancer in mice. Those experiments, by the way, are not exactly   very expensive. You can outsource them. But if you don't have that, you shouldn't   be raising $500 million or even like $50 million.  Maybe $50 million is fine, but not a few hundred. 

But that's the benchmark. But now, coming to the lab, in   2019, we thought about and we started partnering  with companies that generate data automatically   using robotics. We saw that when you've got  still humans in the loop, you would still   be very biased when picking the targets. Any time the human looks at a target, it's   like quantum mechanics. You look at the article,  and it's either a wave or not. Right? [Laughter]  

Or it's a particle. Depending on  whether you looked at it or not, right?  Here, if the big pharmaceutical  executive looks at the target list,   that's it. They already know. They are  bias. They know the targets that they know.   They saw something that is logical, and they  will try to cling to this target choice. 

We actually didn't want to even show humans the  target lists before the targets get validated.   That was the big idea. How do we de-bias  people to allow for more greater exploration,   because every pharmaceutical executive wants  confidence? They actually want to go after   something that is more likely to work within  their short career in the pharmaceutical company.  We decided to build our own lab that would allow  for generative AI exploration from scratch,   not just to train AI on machine-learnable data.  No. That's not what we are trying to do. We wanted  

to allow for genuine AI imagination to take place,  and then you validate those hypotheses that come   from this AI imagination with real experiments  and reinforce those pathways that actually worked.  Let me show you what we did. By the way, during that time, COVID hit.   [Laughter] You probably nowadays would  kind of tend to forget about it, but 2020,   we got COVID and the world decoupled. China went on lockdown, so actually,   inside China, you could still work but you could  not just travel there very easily. I spent 14   weeks in quarantines building my lab. Every time you go there, it's two weeks   quarantine, and I loved it. Again, I think  that right now China is being demonized by  

the entire world, especially by the U.S.  But most of the things you hear on TV   are not true. It's kind of complete garbage. Nowadays, I actually think that fake news is a   real term because people there are extremely  hardworking, they're extremely friendly,   they're very cooperative, and I think that  if Aliens landed in China during COVID,   if they didn't have COVID and they  were friendly, they would be welcome.  I landed in Suzhou and decided to build my lab  there because they have robotics capabilities like   none other. And, of course, the country itself,  internally, they are open. It's just that you need   to spend two weeks in quarantine. I'll show you the footage  

of what happened. This is our company. We  are truly global. We have eight R&D sites.  This is one in Shanghai. That's where we do drug  discovery and many people in drug discovery work   there to supervise many contract research  organizations that make drugs and synthesize   and test them. We have a pretty epic floor,  a great presentation area, super high-tech.  One and a half hours' drive away from Shanghai  is Suzhou (or it's 30 minutes by train).   This is April 2022, so exactly a year ago, we  got this space. We have a logo on the building,   and that is me promising to build  a lab in April 2022, so last year. 

Then again, the COVID pandemic hit, so there  were some restrictions. But as I mentioned,   these people are true heroes. They  worked and slept there in this lab.   I have never seen people more hardworking. Here you have a gas leak and they're still working   in respirators. That's me sleeping in the lab. That is today, so you walk in. We wanted to   make sure that it looks like  a spaceship as well, so people   understand that they are working in the future. It's face activated. You walk in. To your right,  

you've got dimmed windows or where we can un-dim  them and see the robots. But most of the time, we   have some confidential stuff there that we don't  want to show. But we can also show the workflows.  We've got a presentation area. You've got  miniature copies of every room with the robots and   stuff we have. You can actually see the workflow. On December 29th, so just three months ago,   I opened this lab and invited a few partners  from big pharmaceutical companies to see how   it works. This is my co-CEO, Dr. Ren, who is  actually a literal hero of this revolution. 

This is a real workflow from the lab. We  take an animal tissue, send it to the robot,   the robot picks it up, grinds it,  microplates it, does quality control,   passes it to another room. By the way, after that,  the human work is over after we put in the sample.  We've got AGVs, autonomous guided vehicles that  work around the lab to ensure that there are   no human errors. You get imaging, high content  imaging. You've got high-resolution imagining.  In parallel, you start the workflow  for the next-generation sequencing,   so you prepare several libraries. You prepare  the library for whole genome sequencing,   for RNA sequencing, for methylation, and  we also collect a few other data types that   I don't want to talk about because people will  say that they also have it (even if they don't).  We prepare those libraries, give it to  the sequencer. We get methylation data,  

transcriptomic data, and a few other data types  that we feed into AI in addition to sequencing.   Again, this AI has been validated and  we know that it can discover targets.  Now it starts the exploration phase.  It picks the targets. It looks at those   that already have compounds and picks  those compounds from the compound hotel,   puts those compounds onto the liquid handler. here they get micro-plated, they get aliquoted,   and being prepared for what's yet to come. We also do a bunch of quality control experiments.  

Here we can do also Echo, some enzymatic assays. In parallel, you pick up the samples from the   incubator that you put in there originally,  incubate them with the predicted compounds,   put them back. Put the compounds in. Put  them back in the incubator. After that,   you have three parallel workflows. Again, you get high-content,   high-resolution imaging. You get methylation,  transcriptomics, and a few other data types.  AI learns if it picked the right targets, if those  compounds worked and did what you wanted them to   do. For the most promising targets, humans would  get the signal, and they can also do human-level   validation, and we can pursue some of  those targets. But most of the time,  

the AI just picks those compounds  and trains to get the right targets.  In parallel, we have a CRISPR workflow. For those  targets that don't have the compound, we can also   do CRISPR screens. But originally, you actually  want to find those targets that are addressable   with small molecules, and it learns all the time. Generative AI is being reinforced using real   experiments. I don't think that there is  anything like that, so some people talk about  

using machine-learnable data for training.  This is not what we do. We already trained,   so those are pre-trained models running the lab. Alex, we're just about out of time. Let's   take one last question quickly from  Twitter. This is again from Arsalan  

Khan who wants to know what happens when  the underlying data is bias. For example,   let's say that you don't have sufficient  data gathered on the effects of a certain   medication or on certain groups of people  since maybe those groups could not afford   the medication. How do you handle that situation? In our case, we can actually pursue many, many,   many alternatives. Usually, you start with the  alternatives and the pathways where you do have  

the data. Again, getting one drug to the market,  its traditional approach is $2 billion. Our   approach is going to be significantly cheaper and  faster, but still, it costs you a lot of money.  You have very, very shots on goal, so usually,  you select the cases where you do have the data.   If you don't have the data, you need to  generate the data either experimentally   or use a generative approach. But then the  probability of success is going to be lower.  Again, drug discovery is a brutal  thing. Right now, this year,   we're going to see maybe 60 million die of  aging and other diseases because there are   no cures. By the way, there is aging anyway,  so if diseases do not kill us, aging will. 

Life is not fair, right? Whatever you do,  you still keep losing after a certain age.  You need to start solving problems in the order  of priority. The priority is A) demonstrate the   clear case that AI can be used to get  the drug through human clinical trials   discovered from scratch with a novel target.  That takes a long time and a lot of capital.  Before you complete that, you shouldn't  be thinking about other outlying cases and   significant democratization of this because it's  going to continue to be a very expensive process.  

The way we want to democratize this,  and I'm going to show you a slide   just so you understand this is not  just words, we do have plans for that.  The idea that I currently carry is to validate  this lab as much as possible with my own projects   and also customers' projects. But then  miniaturize this lab to the level where   I can maybe make it into even two rooms  or maybe even one room if I am lucky.  We want to expand the lab and add additional  capabilities, but we also want to miniaturize   the lab, optimize it so that it becomes  small enough. This lab, I would need to   build in 3D, so it has to go to  the ceiling of the laboratory,   and humans would not be able to walk in,  so that's very dangerous because very   often you need repairs or reagent changes or  something breaks. So, you need to be very good   at this to miniaturize this kind of technology. My idea is to miniaturize and put it in hospitals  

so that the hospital would acquire all  the capabilities that my company has,   so they don't need to share the data with  anybody. I don't need their data. That's what   people misunderstand usually about my company. We have a policy. We don't touch your data.   I don't want your data. We actually  will refuse your data most of the time   because it has dangers in there, including  all kinds of geopolitical dangers.  We want you to be able to discover targets. If  I give you the lab and the software to run it,   you can discover targets at hospitals' premises. What do doctors do well? They can take biopsies.  

They can throw biopsies in the lab, and the  lab would help them identify a pathway to   treat the patients better now. But also, as you  get more samples, you can identify new targets.   That's the way to democratize drug discovery  globally for the first time in human history.  If you can do drug discovery at hospital  premises, run by physicians and physicians   are usually not at the level (unless you are  talking about physicians who work in pharma,   drug discovery). Physicians don't have the  capabilities usually to discover drugs. But with  

AI and this robotics capability, physicians will  have the ability to acquire those superpowers.  Imagine the countries that never discovered a  target that resulted in a drug. Naturally, that's   most of the countries. You put a few of those  robots in hospitals and now countries like Saudi   Arabia or the UAE (oil-rich countries), they can  now convert their petrodollars into viable drugs.   How cool is that? And you can put it in Africa. We have a couple of other questions. I'm just   going to ask you to answer them very  briefly because we are past time. 

this first one is from Jedd McKensie who  says, "What is your opinion on the use of   AI in medical writing and regulatory  submissions?" Very quickly, please.  In medical writing, currently the accuracy  of generative systems is very low. Before   you have massive benchmarking and validation of  medical writing AI, you should probably not use   it. By the way, if you get caught doing  that later on, you might get prosecuted.  To just demonstrate how it works, I published  a paper with ChatGPT recently and made a few   cases around that. You can look at my article  in Nature Medicine about the dangers of using   generative AI for medical writing. Specifically regulatory submissions,  

no way. You don't want to screw your most  important time of your life when you are   doing anything with the FDA. That has to be  pristine. IT has to be super-regulated. Double,   triple quality controlled, and you  want to put more effort into doing so.  However, in parallel, you can actually  do some benchmarking with AI just for   internal experiments. But you shouldn't  be using this for submission. Yeah. 

Another question, again very quickly – it's  a complicated question – with generative AI,   do you run the risk of violating IP because of  the publicly available data that you're using?  It depends. First of all, if the data is  publicly available and generated by the NIH,   it is generated with a purpose of exactly  this. You use it for experimentation and the   government doesn't have any IP in that. Usually, this data is available,  

so Google is doing it for you right  now. They are helping you with search.  You've been using other people's data  to make discoveries all the time. You   are a generative system. Anybody who  is creative, they are also generative.  [Laughter] I try to nowadays use creativity as   generativity. You can substitute those terms. There will be IP issues when you are   utilizing proprietary data. Somebody's proprietary  data, if you get the full-text articles,  

for example, without paying for them to the  publisher, then it is possible to watermark and   notice that this data was used for the generation  of your content, the company that is doing this,   it might get prosecuted. There's a long debate about   this because, in this case, AI is just like  a human. It needs to be treated as a human.  If the human saw something and learned and  then generated something new, the level of   novelty is very high and there is no way to  trace where you got the original, you know,   maybe pre-training data for yourself. Then there  should not be IP issues. But if there is a clear  

trace, then yeah, there will be IP issues. In chemistry, this shouldn't be a problem   because there you rely on massive chemistry data  sets. If your molecule is not similar to anything   that's published or patented or in the process  of being patented, it's very diverse, then you   should not have any IP issues. You own the asset. With that, I'm afraid we are out of time. Alex,   thank you so much for taking your time to  be with us today. I really appreciate it. 

Happy to be with you. Again, I think AI  and robotics are going to do great things.  A big goal for all of us is to solve aging. It's  a big statement to make, but I think that once   you set this goal high enough and the bar high  enough, everything else starts looking achievable.   I think even aging is achievable, but that's  where AI is going to make the biggest difference.  Actually, my biggest contribution, I think, in  generative AI was starting to train generative   systems on longitudinal data to generate  synthetic data with age as a generation   condition. That allows you to play with much  more data, synthetic data, than you can think of. 

I think aging should be a priority for all of us. Absolutely. There's no doubt that AI is going to   have a profound impact on all parts of our  lives including drug discovery and medicine.  Thank you, everybody, for watching. Now, before  you go, please subscribe to our YouTube channel   and hit the subscribe button at the top of  our website so we can send you our newsletter. 

With that, thanks for watching, everybody. I  hope you have a great day. Check out,   and we will see you again  next week. Have a great one.

2023-04-18 02:31

Show Video

Other news