243: Why AI Still Hasn't Revolutionized Drug Discovery (Yet) | Thibault Geoui, PhD

If AI is already being used across the drug development pipeline, why hasn’t its impact matched the investment?

AI can help researchers review scientific literature, predict protein structures, prioritize molecules, assess toxicity, support clinical trials, and monitor adverse events. But access to better tools doesn’t automatically create better drugs.

In this episode, I speak with Thibault Geoui, Science CDO and host of the Tech & Drugs Podcast, about where AI is making a practical difference in drug discovery and development—and where the results remain limited. 

We map AI across the full drug development funnel, from basic research and target identification to preclinical testing, clinical trials, regulatory documentation, commercialization, and pharmacovigilance.

We also discuss why digital-native tech-bio companies may be better positioned to benefit from AI than traditional pharmaceutical organizations. The difference isn’t simply the model. It’s how data, people, laboratory experiments, and AI tools are connected inside the workflow.

For digital pathology professionals, the conversation becomes especially relevant when we examine AI-powered biomarker development, the role of pathology in pharmaceutical research, and the Roche–PathAI case discussed in the episode.

And, of course, we talk about the problem every AI user eventually faces: an answer can look polished, specific, and completely convincing—and still be wrong.

Episode Highlights

00:00 — When convincing AI output creates more work
Why AI can accelerate information generation while increasing the time required for review and verification.

02:15 — From structural biology to science and technology leadership
Thibault shares his background in X-ray crystallography, structural biology, scientific data, and digital product development.

15:36 — Understanding the drug discovery and development funnel
How thousands of potential compounds are narrowed down through discovery, preclinical research, clinical trials, and approval.

20:00 — AI for scientific literature review
How alerts, filtering, summarization, and information extraction can help researchers manage a rapidly growing scientific literature base.

22:32 — AlphaFold and protein structure prediction
What faster access to predicted protein structures changes for researchers—and why structural prediction alone doesn’t solve drug discovery.

24:13 — Searching an enormous chemical space
How AI can help design and prioritize potential molecules for synthesis and experimental testing.

25:50 — Predicting efficacy and toxicity
Where AI supports preclinical research, why the models remain imperfect, and why experimental validation still matters.

29:38 — Has AI changed drug development outcomes yet?
A practical discussion about drug approval rates, AI investment, uneven returns, and the difference between deploying a tool and integrating it into a process.

34:33 — Why traditional pharma struggles to scale AI
Siloed data, legacy systems, organizational complexity, and the need to build reusable data workflows.

37:57 — The “lab in the loop” model
How tech-bio companies connect AI predictions with wet-lab experiments and feed the new data back into their models.

44:37 — Can tech-bio companies shorten development timelines?
How digital-native organizations are changing parts of the discovery and preclinical process.

58:00 — AI, pharma, and digital pathology
What the Roche–PathAI case discussed in the episode may indicate about the role of pathology data, biomarker discovery, and pharmaceutical workflows.

01:06:17 — AI errors in regulated environments
Why responsibility remains with the person or company submitting AI-assisted work, regardless of which tool produced it.

01:17:37 — The growing cost of AI tools
Subscriptions, token limits, model selection, AI orchestrators, and the need to use expensive tools more intentionally.

01:27:50 — What successful AI adoption requires
Starting with focused pilots, training scientists and technologists together, and treating implementation as organizational change.

01:30:26 — The AI quirks that still frustrate users
Hallucinated information, ignored writing instructions, stylistic habits, and poor awareness of time and context.

The episode’s timestamped themes and examples are documented in the supplied summary. The broader discussion covers AI from literature mining and molecular design through clinical development and post-market monitoring. 

Resources Mentioned

•  Thibault Geoui’s LinkedIn profile 
• Tech & Drugs Podcast
• MIT NANDA study on generative AI implementation and return on investment 
• Insilico Medicine as an example of a digital-native tech-bio company 

AI is already changing how scientific work gets done. The bigger question is whether organizations can redesign their workflows, train their teams, and maintain the human oversight needed to use it well.

Listen to the full episode for a practical look at AI in drug discovery, drug development, and digital pathology.

Transcript

00:00:00

The problem with AI is that it tends to do thing extremely convincingly. So the other day I used the eye to create a list and then I looked at it quickly and I was like a it looks great and I sent it to a colleague and I say hey look at the list. I mean it's really nice. I mean it was able to pick up new stuff and then you know I sent it to him by email and then I started to read more of the list and I'm like just did it by email but it was well done and so convincing and I think this is the the

00:00:24

 biggest problem with a lot of AI at the moment is that it still elucidate quite a lot in very convincing manner where it accelerate your workflow it's in the production of information where it's slow your workflow is that you need to spend more time reviewing Welcome digital pathology trailblazers. Today's episode is going to be about AI in drug development. AI in yeah drug discovery and drug development. AI for drugs. And my guest is my former colleague Tibbo Jawi. Welcome Tibo. How are you doing today?

00:01:04

 >> Uh I'm I'm very good. Thanks Alex for having me. And Tibo, you are also a podcast host. Your podcast is tech and drugs and you have interviewed pretty prominent people in the pharma space like um individuals from GKO bioworks, Tim Hawker and Dove Gerts who are everybody who googles them is going to see what kind of names these are. So because we before we start with you and I'm going to ask you about your background in a second. I'm just going to set the stage. So we I'm still at Tr

00:01:40

 River Laboratories and this is where we were colleagues for some time and we cover drug development. I would say we cover it from um a little bit of a different angle. I'm very much focused on the digital pathology and how digital pathology can enable that. you cover more broader the AI angle. Um, so let's start with you and give us a little bit uh about your background. >> Okay, I'll I I'll try I'll try to be to be short so you don't like stop me if I start to to talk for too long. But um I

00:02:15

 like to talk um >> you have a you have a cool story so go ahead and tell us the story. >> So I'm I'm you know I'm a scientist by training I guess like like most people in in our you know in our circle. What kind of scientist? >> I I'm a structural biologist. So mix of biochemistry, biohysics, done my PhD in so on the X-ray crystalography of the Epstein bar virus of some some proteins. So uh when I did my PhD, we had um structural proteomics program. So we were like taking all the proteins from

00:02:50

 the from the bar virus and trying to to solve as many as possible. So I worked I did my PhD in France in in Grenobal in my it was such a cool lab because in my in my backyard we had a synretron. So I don't know if you're familiar with synretron but you know it's like >> I don't even know what that is. >> What's a synretron? >> It's the biggest scientific instrument in the world say for particle accelerator that you have in Geneva that is like >> that's this CERN thing that sends

00:03:20

 electrons. It's like it's like CERN in smaller like CERN is 27 kilometer. Mine was just 1 kilometer. >> Um >> so you had like the baby version of it. >> Yeah, I had the baby version and the and the CERN it's accelerating particles. Uh and a synoptron it's different. It's you know you have electrons and then when the electrons are are taking a turn uh they emit X-rays. So it's a bit like think about the X-rays when you go to the doctor where you have like a small

00:03:50

 X-ray machine. This one is like a billion times stronger and and you can do all sort of really cool thing with that and and essentially the the way we describe it is it's like a giant microscope. Uh but the problem is that you know the the smaller you want to go the bigger the instrument. So you know when you when you have a microscope you know you can you can see pretty tiny stuff. I mean with electron micros actually can see really really tiny stuff. But at least when I was doing my PhD, the only way to see like you know

00:04:19

 uh things at the angstrom level so like a tenth of a nanometer was to use uh was to use X-ray crystalography. So you know with X-ray crystalallography you can have like a an atomic resolution structure of of a protein. Why do you want to do that? You can ask me. Well, you know, if you if you know what something looks like, then you know, you can start to understand, you know, like a link between the the shape and the function, the structure and the functions, you will you will identify uh you know, cavities in the protein where

00:04:50

 you can design specifically design the drug and uh and and just you know like something not scientific at all, it just looks beautiful. You know, [laughter] one of the thing that drew me to that is that I really like because, you know, sometime you work on, you know, you when you do biology, you know, most of the time you'll have a picture of a gel gel with like with a band, you know, it's black and white. >> I did PCR for my PhD. >> Yeah. Yeah. So, PCR [laughter] Yeah. or or SDS page, you know, and you get lots

00:05:21

 of lines and then you're super happy when one line is like a little bit below the other one. It's a bit boring, you know, when you get a protein structure. It's beautiful. You have lots of colors. You can animate them. So, you know, this part is not the scientific part, but I I was kind of drawn to uh the beauty of it. And and the other thing is that when you enter a synretron, it's like being in a in a spaceship. It's crazy. It's like Yeah, it's like you were in Star

00:05:47

 Trek. So, I think, you know, like when I when I was a kid, I wanted to uh to be a cosmote. like like a lot of people or a lot of people who turned out to be scientists you know who were actually I'm quite claustrophobic so I was discussing with my with my wife today >> yeah I was discussing with my wife today we were talking about you know what kid wants to do in the future I said yeah kid want to be astronaut technician like claustrophobic so I'm like yeah uh so then I went and did did something else

00:06:15

 before my PhD I was convinced I wanted to be like a a scientist for the rest of my life I wanted to be a professor more and during my PhD I was teaching you know I had like a part-time uh teaching assistant job and uh and after three years of doing you know intense research and intense teaching I realized I didn't really want to be a professor anymore and although I still I love science >> that's okay >> yeah know I know I know but but you know like I think there is a bit of social

00:06:43

 pressure that you do a PhD you get to do post and then and then going out of science it's like >> I don't I should not say podcast. But >> yeah, but it's not only that. I mean, okay, it's a public podcast, so it should be on record, but it's a bit like a mafia. You know, you're in a family, you can't leave the family when you eat. >> So, [snorts] >> I know. >> So, so yeah, I did I did, but I still wanted to stay in a scientific field. So, I went to work for companies that

00:07:09

 are, >> you know, the same way. I never wanted to be in academia >> with like all respect to the research that they're doing because of the dynamics and it felt such a political environment to me and then I realized that like any organization is going to be a political environment to one extent or another. This is how like people self-organize. But this was my my main driver not to do academia because the currency are your papers. And I'm like that's invented currency like citations

00:07:45

 and you do bend >> over backwards or whatever the saying is to like publish and anyway uh you know that's a philosophical discussion but I had that feeling as well when I did my PhD. I thought no >> I think many people >> yeah I think many people you know like out of all the people I was uh you know in my in my master and then PhD I don't know maybe 10% or 5% ended up actually getting you know a tenure track uh a lot of them you know went through you know like three free of dogs and then then at

00:08:20

 some point they either got a job or you know an academic job or actually >> left for the industry and then you know I think I realized quite quickly Look, I don't want to follow this path because I don't want to leave because I I can't do it. I want to live on my own term. But then it's not, you know, doing something by default, but I I was like, no, I want to do something that I would really like and also explore something, you know, what doing my PhD. I had no idea what people with a PhD in structural biology

00:08:49

 on viruses can do in in in outside of academia. So I my first job that I landed was um was with a company called Kaen. So people in the lab with no ken you know like they do consumable instruments uh diagnostics all sort of stuff for scientists. I worked with their product. I like them. I was like hey I like their product. I know their product. So why not working with it? So I worked with them for five years. I was a product manager and then um I learned tons of stuff. Uh I really liked it. But

00:09:19

 then someone one day contacted me a recruiter and say hey you know I had this job at health publish. My first reaction was like I don't want to work for sure and they were like hey wait a moment you know it's it's yeah it's for LC but it's a it's in the life science data and analytics team it's about data and that was in if I recall in 2011 so 2011 you know was quite a long time before we started to that time we were talking about big data I didn't really know anything about that but I was like

00:09:49

 hey it sounds and I think it's a sh I didn't actually know why but uh but then I went there and actually said 12 years I I've done so many things. I started in marketing then moved in digital product management in uh strategy and innovation. So essentially what we were doing there is taking uh data extracting data from unstructured content. So what is unstructured content? Essentially, you take a publication and you extract the data. In the publication, you might have something that says protein A binds

00:10:19

 to protein B and uh has a role in a certain disease, but then it's in a text, you know, it's all over the place. So, you take that and you put it in the in a database and then you you you build a search engine on top of it and depending on who the search engine is for. If it's for a chemist, you will you will design it in such a way that you can glow molecule because chemist like to glow stuff. If it's for biologist and you need to have text and you need to make it simple because they don't like

00:10:44

 to use operator and things like that sometime they don't really know how it works. If it's for pharmacologist you know you will have another way of of designing it. So it was really cool because it was still super scientific. It was in a in a field where for some reason I had this intuition that it could become important and then over time it became more and more important because we started with more search engine then we started to create a data product where people could use the data

00:11:12

 to actually train model at some point we we the very beginning of the GPTs and the LMS we had a project where you know we worked with with um AI studio to create our own AI AI products And um one of the thing that I liked is that it touched on so many areas of science and drug discovery and development and I found it fascinating and for me that was you know one of these happy accidents in a career where you you decide to to take something where you're not really sure if it's the the right thing to do but it

00:11:45

 ended up being really >> exciting for me because >> yeah because then that's how I started to be in the data in the field of data which I think now is definitely the I mean for me that's the you know the the best place to be in our industry after um after LCV actually that's when we work together you know I was at Charles River and Charles River was you know for in many ways my dream job because I you know at the time I reported to to the CIO with a ti to the cso and the job was

00:12:14

 really to look across all of research and say look at you know how can we use technology to accelerate to accelerate what what we did at Google >> unfortunately didn't last for very long Because I think I I arrived in the company when uh you know like the not only the company but the the industry as a whole was going through a bit of turmoil and at some point there was a a reduction in force and uh I lost my job which was something that I was so fed off that it will that will fix in my

00:12:45

 life I lost a job but I mean that's a discussion we had separately I I you know I had done a lot of work online you know I was kind of blogging writing on LinkedIn daily for for a couple of years already. So I had quite a lot of visibility in the industry and then um you know for for some time I I uh I started my uh my own business. I started I started the podcast as you said taking plugs then I got a few client and one of his client is the company that I'm working with ask me okay can you want to

00:13:16

 join and start practice uh and now I'm working for an innovation and engineering company again something very different but really cool because again you know you touch on so many different aspect of our industry it's very technology and innovation focused and it's really about solving interesting and important problem and it's still you know like it's still very much focused on how can we help organization you know work faster better be more more efficient more innovative you know using

00:13:48

 this mix of uh of technology and science and and for the essentially for the past 15 years I've been at this intersection of how can technology helps with uh doing better science and that's you know regardless what I will do in the future that's that that's a place where I want to to work it's intersection of technology and science because I think that you know when you bring those two things together especially now and we talk about it we talk about AI I think this is the the most interesting uh

00:14:19

 thing that ever happened to uh to our industry >> so that's my >> and you're like >> that was I think the longest intro I had on the podcast but we keep it >> sorry sorry >> no but no because it's I love it because it shows like how many different paths can a scientist take. Uh and also that then scientists from different angles, different areas meet how we met at the same job, right? And the data and technology theme also very shows very much in your uh LinkedIn content. So whoever is not

00:14:58

 following you h I'm going to leave the link to follow you on LinkedIn. I love the insights uh that you give from publications from uh what's happening with AI in drug discovery and development. Um I want to bring us back to this topic. Let me go through what drug discovery and development funnel looks like because a lot of people think of drug discovery, drug development. when they think of drug discovery and drug development, they think about the last stages, the clinical research and then approved drugs that a doctor can

00:15:36

 prescribe, right? And there's a lot more before that. So when you look at the drug discovery and development funnel and the number of candidate compounds, so like potential molecules that could become a drug, you start with the basic research and then and the numbers vary depending like what kind of uh image you Google and what kind of source uh you look at. But let's say the basic research uh we're going to have 20,000 compounds. Then uh we go to stage two, drug discovery. we're gonna uh reduce it

00:16:10

 to 5,000 compounds. It can be more or less and then only uh we're going to go to pre-clinical research which is where I sit at the contract researchers organization where I work at Charles River Laboratories. This is where you do work on animals. Only then out of those 250 compounds that made it to the preclinical research because all the rest got discarded. You take five to clinical research and out of those five maybe you get one approved. These are all the different levels where you can apply AI and what's your take on that?

00:16:48

 Where is it being currently applied? where is it not being currently applied and let's start the discussion on AI and drug discovery and development. >> So, so I think it this picture almost misses one uh you know one box at the very top um because you know before before you even have like compound and everything you know there is a lot of fundamental research that happens even before entering into pharma. >> So I'll take I'll take a popular example now. If you look at the the weight loss

00:17:20

 drugs, you know, like the GLPS. >> Yes. GLP one, >> you know, like like the typical thing that you will read online is that developing a drugs takes 10 to 12 years. It will c it will fail 95% of the time and then it will cost two billions. So >> let's keep those numbers in mind. >> That is always the why are drugs so uh expensive? Oh, because the research and development costs so much. So they have to recover the cost. This is like the standard answer why it's so much.

00:17:49

 >> Yeah, >> we can agree or disagree and it's multifaceted but yeah. Mhm. >> But even this picture I think is uh now is evolving because you know there was a paper recently from the same offer that gave the two billion cost and there was a recent paper when they updated it they said hey actually if you look at the largest 16 or 17 pharma company it's not two billion on average but it's six billion. So, >> oh okay, >> the large actually the larger the farmer

00:18:17

 company. >> Yeah, but that's the thing which is super interesting. The larger the farmer company actually the the least efficient they are and the the more costly they are. So that's another thing we can we can talk about. And then if you look at the very top actually you know like the basics that happens in academia. So if you look at um I I'm sure I'll get the date wrong but I'll try to get at least the ballpark figure. the the very first work that eventually led to the

00:18:44

 development of the the GLP DBTR. So the the weight of drugs drugs for uh against um diabetes started in the 70s. So it was essentially finding okay something which is involved in the in mice in regulating appetite you know started in the 70s then they identified some receptors and some you know some peptides in the 80s and then I think in the 90s they're like oh okay maybe this this thing could be druggable and then at some point you know that landed into uh into pharma but there was like 30

00:19:19

 years of basic research done in academia before this got into a stage where a farmer would like hey we can make a drug from that and then it took like 15 years to actually take this and turn it into a into a into a drug. So now where is it that uh AI can be used or is being used I would say everywhere in in this pyramid at you know at different level of intensity and with different level of success. I think a lot of um they are they are they are quite uh you know like no-brainer use case you know a lot of a

00:20:00

 lot of uh time that researchers spend is you know trying to find information like going through if you remember during COVID I think in in one and a half year there were something like 50,000 publication published on COVID a lot of it was absolute garbage but then you know like if you if you have a Google alert saying hey tell me when whenever there is a navigation then like every day you get hundreds how do you read that how you know just reading the the title not even the title and abstract the title of 50,000 stuff takes a lot of

00:20:31

 time so you know if you can have AI helping you you know like filter things and tell you okay this is important this is not and then actually summarize thing extract some information that's already like a huge amount of time say >> and then there of >> so uh this is something actually that it doesn't have to be in drug development. Anybody can leverage that right now. So I have an alert obviously for digital pathology and AI and I have like specific rules in perplexity. Okay, it

00:21:03

 has to be relevant because this and that. So that's something that uh that people can uh my trailblazers who are listening to it right now can implement >> and for themselves you know now you know you can you can be like you can start to be really sophisticated with with those tools. I mean complexity is great if you use clone. I mean you can even build agent and uh and you know like connect things with different tools and uh and tell me you know like if I have uh search this side that if uh if things

00:21:34

 meet some criteria that send me an alert and then you know you can even write a summary. I mean you know like I've seen people building really really interesting stuff with with agent and and that's the thing you don't have to be a coder you know all you can you can do things you know you can just describe you know what you want and then it will it will be you will be able to pretty successfully get get meaningful information so I think this is no-brainer this is something that works

00:21:57

 really well that is being used now now if you if you move uh more in the in the early discovery you know like identifying targets designing drugs so these are some of the popular popular example of of using AI um you know like combing through a lot of experimental data doing analysis. So you know I said I did my PhD in structural biology. So it took me four years to to solve the structure of a protein. And in 2024 Demi Asabis and John Jumper from deep mind they released alpha fold. And alpha fold

00:22:32

 is a model that uh can take the sequence of a protein and and give a structure which is you know almost indistinguishable from what you will do you will have from experimental data. So of course there is always value in doing the experiment but you know in a in a few minutes you get something which is almost as good and completely usable in the in the context of of research. Of course there are tons of caveat you know it doesn't necessarily work for everything but I think the important thing is that it took us 50 years to get

00:23:05

 170,000 protein structure using experiment and now we have 200 or 300 million that we've done since Alpha was uh was released. So he's been an incredible tool in in research. So it hasn't solved drug discovery and development because drug discovery and development is not just having a structure but it's a very important piece where people used to spend a lot of time. So, so there is this then you know we have tons of tools now once you have a structure to also design like a molecule and you know and and propose a

00:23:36

 molecule because if you look at so I'm not going to talk about antibodies but if you just think about the the chemical space of drug like small molecule so what what what do I mean is you know a chemical molecule that has the characteristic of a drug there is 10 to the power of 60 drug like molecule. So that's a chemical space. So just 10 to the^ 60 is so big that it's difficult to even comprehend how big it is. So just to give you two brackets, if you were to count every grain of sand on Earth, it's 10 to the^

00:24:13

 of 25. Uh and every single atom in the universe is 10 to the^ of 80. So 10 to the^ of 60 is a crazy big number. >> That's a lot. >> So my goodness, let's be clear. There is no way on earth that you can synthesize 10 to the^ 60 molecule and try all of them. You can't even compute them. I mean it's so big that even in a computer you can't represent all of them. But the number of molecule that you can design in a computer is several order of magnitude greater than what you can

00:24:45

 actually make you know like physically. So if you can synthesize them and then you have the shape of your of your target then you can you know like you can design uh in silicon a number of of potential molecules. Then if you continue you know like down the line so now you have a molecule or several molecules you can start to predict okay will those molecule be demonstrate efficacy. So efficacy is does it do what it's supposed to do. But if you have something that is supposed to uh you know uh cure you know trigger

00:25:18

 weight loss or cure a headache if it does it then it's it demonstrates efficacy that's one thing and then the other thing that uh Charles goes a lot is is it toxic >> safe >> yeah is it safe so yeah is it you know safe meaning not toxic so if I give it to you will you have a heart problem you know will you cure your headache but then you have a heart attack well that's not a I mean it could be so efficacy but then it's toxic. You don't want that. So again here there are plenty of AI models

00:25:50

 where you can have prediction it's not yet perfect. Let's be clear. I mean uh for for a lot of those things um you know people are like you know what I'm not going to take a drug that was just tested in silico. I still want uh with everything we say around reducing animal experiment. You know we are still at this stage where you know we still want to check uh in uh in animal enables and v with what we call NAS like alternative alternative ways of testing. >> Yeah advanced methodology whether those

00:26:25

 uh those molecules are are safe but then in silicon in silicon methods are playing you know an ever greater role. Then you know when you move to the clinical phase in clinics AI is starting to play a bigger bigger and bigger role. I think it's not as advanced or developed as the the discovery and critical phase but there are plenty of use case in in clinics like for instance how do I select uh which patient is the right target patient you know what are the inclusion exclusion criteria >> uh and then a lot of clinical phase are

00:27:00

 you know almost like a supply chain problem you know like how do I send things at the right place how do I select the right hospital or select the the right principal investigator that you know has a good track record of doing things all those things then you know you think about I mean be in the preclinical phase and in the clinical phase there is tons of regulatory documentation and here AI is used a lot you know disc they made it public months ago but they they were able to reduce you know from several weeks maybe five

00:27:32

 weeks to you know couple of days uh you know the time it took to have first good draft um of of regulatory regulatory docu documentation using AI. So that's also something where it's used and then of course you know like after that there are the commercial phase and uh the pharmacco vigilant phase you know once commercial phase it's it's pretty obvious what it is it's setting drug identifying the right the right patient population the right GPS the right commercial strategy but that's something

00:28:04

 that just I mean quite similar to what happens in many industries so I think here a lot of AI that has been developed for commercializing any asset can be reused in pharma One thing though that is quite specific to the pharmaceutical industry, the pharmacco vigilance phase which is once a drug is on the market since you give it to a much bigger population than what was in the clinical phase of course you will have uh some adverse event that you couldn't catch in the clinical phase. Clinical phase might

00:28:33

 have several thousand uh people taking a drug. So you know like there is but imagine there is a serious side effect that happens only in one person out of a million then you will not be able to catch it. So pharmacco vigilance essentially it's when once a drug is in the general population if there is a problem you need to be able to detect it but detecting it is quite challenging because it's being reported in many ways either publication or doctors will have you know if they see a patient taking

00:29:03

 the drug and they notice a specific event they will have to report it here you can use AI also to uh to detect signals >> so yeah I think you know of course I haven't been exhaustive but there are so many areas has now where AI is being used. But you know the the challenge remained you know like we still um if you look at the FDA approval you know every year on average you have like maybe 45 or 50 drugs approved by the FDA we haven't seen yet an impact on of AI on the approval. You know one thing one

00:29:38

 one number I always like to to use to put things in perspective is that there approximately 24,000 known diseases 14,000 are rare disease which mean in the US a rare disease is when you have one person out of 200,000 which has the disease we have something like 4,000 approved drugs and of course 4,000 drugs approved drug doesn't mean against 4,000 disease because you have it 10 drugs sometime for the same disease. So maybe it's for a thousand or thousand disease and then the the drugs that are really

00:30:09

 curated meaning you take the drug you're cured maybe like 200. So you know at the rate where at the rate at which we are proving drugs today it will take us another 500 years till we cure all diseases. Uh that that's kind of where we are. >> We still have a job for some time. If you're working in drug development there is still still a job. M so you know what one one theme uh you know you you cited or or you explained an application in each phase of the funnel uh the the theme that keeps coming up is scaling

00:30:46

 scaling what we would do manually right scaling the work on protein crystals scaling logistics of a clinical trial but I don't know are we like successful with this scaling if There's okay we still have the same number of approvals but does FDA get more submissions let's maybe talk about a success case is there somebody who does this well and I remember wait uh you posted a you posted a LinkedIn post about 95% of people not seeing any return on investment on using AI and then 5% actually seeing exponential

00:31:35

 crazy returns on investment because they're doing it right. So it's not that AI is not working, it's that we're not working well with AI. So who's working well with AI in the drug development space? >> So So first of all, the the paper that you're citing, it's uh it's called the MIT Nandanda study. uh I guess you can put the the link uh and it was not specifically for the pharmaceutical industry but it was really about generative AI not even you know like all

00:32:02

 type of AI because you know we tend to we tend to think about AI we think about LLM and generative AI but uh it's more than that but it was really the case of generative AI and companies implementing generative AI and they they were like oh you know like we bought we we bought a license to to uh chat GPD gave it to all of our employee we like yeah sure I mean you know it's it's not because you give a tool to everyone but they will know how to use it and it's it's well implemented in their your

00:32:32

 processes and so yeah g you know like give me all the tool of a car mechanic I won't be able to fix my car you know I think because I don't know how to do it uh and I think it's the same with uh with um with with AI I was talking to a company founder recently and he and he was telling me something which was interesting. He said, you know, uh we were talking about agents kind of next stage of of the most modern way of of using of using AI. And he said company founders that are like agentic native

00:33:05

 like the most you know recent founders that are really deep into a think completely differently as you know normal company founder. Um and and what he meant by that is that if you look at uh imagine you're like a senior executive with 35 years of of leadership experience. you're being called to lead a new company. So you will think about your company structure. So you know how many people you need marketing, how many people you need in research, how how to get people to interact together. And he

00:33:34

 said in the agentic world, I mean forget about that doesn't work that way. And uh he was giving me an example that you know they wanted to hire someone in marketing uh and they were initially thinking you know instead of having a team of 40 we'll have just you know like one person and you will have a big budget to manage agents and he said well we kind of pushed it even further we decided not to have this person and just have agents managing this uh and he was telling me um every morning when I wake

00:34:04

 up I'm thinking what is it that I'm doing that I should not be doing and that I can automate and That's kind of you know like if you think about it that way well I think these are the company that have disproportional return because they they think completely differently they don't think about a typical pyramidal structure it's it's it's like a a web network of uh people working with agents augmenting them and so now how does this apply to uh drug discovery and development so even if you think

00:34:33

 about your typical pharmaceutical company I mean it's it's very structured it's very rigid There are lots of lots of silos, you know, like the data that you generate somewhere, people don't even know >> that exist in the same company and they go and buy it the data somewhere else from another company because they don't know they have it. >> Mhm. >> Exactly. >> Familiar with that. >> Yeah. And I mean, you know, like I I wouldn't blame [clears throat] the

00:34:59

 traditional pharma company. I mean, some of them have been around for hundreds of years and imagine the complexity. I mean when I when I look at my own computer after four years of using it and how messy are the folder and the file >> I need an agent to clean it up for me. >> Yeah. But you compound that by 100,000 employees and then do that over you know like 50 years or hundreds of years. And of course then you don't even have just computer you have like printed files and all those things. There is no way that a

00:35:26

 traditional company can leverage AI in the same way that a company that started digital like you know what we call tech bio I mean company that started digital they knew from the get-go that they need to produce data in a way that data is reusable >> you know like the the I would say the old way of thinking about research is very linear is I have a scientific problem I'm thinking about my experiment designing it I run my experiment do the analysis produce data I analy that I answer my question and then I move on to

00:35:58

 the next question and then my data I keep them somewhere. >> That's it. >> Yeah. >> But then if you think about it in a in an AI way is that the data that I produce will be used by someone else that I don't know to train a model. So it needs to be produced in such a way that it's kind of a you're creating a virtual circle like you're creating a learning organization. To be able to create a learning organization you need to produce data in a completely different way. People need to think in a

00:36:23

 different way. So you know anam an example that I always like to take is a company like incilico medicine they've been very successful at at applying that uh you know at the pre pre discovery and preclinical phase and now let's talk about what do they do they're a drug company right a pharmaceutical company >> it's it's the poster child of a tech bio um so you know like so you have pharmaceutical company then you know the next wave of innovation was biotech and Now that's what we have we call tech bio

00:36:55

 you know it's kind of taking things and putting tech instead of bio >> to to power bio >> exactly uh and uh and the the way they do thing is that a lot of them when they started they didn't even have a lab you know they were like we can have data we can get data from somewhere from public sources or you know working with a CRO that produce data for for us then what we do is we develop models >> and then you know we will take the data you know like do some prediction then get a lab to do the next experiment

00:37:26

 for us. feed the data back in our model and then make some adjustment based on you know other real live data matching what we predicted or if they are not then it means that something was missing in our own model and we can learn and we can improve and then over time those companies started to be like well you know I want to have my own lab but then how do I design a lab in a modern way but a lot of them they have autonomous lab or semi-autonomous lab so there is a lot of robotics and the most modern uh

00:37:57

 the most modern iteration of that is well I can get the models working directly with uh with a lab it's kind of a digital physical bridge uh because I actually wrote about that but you know you can think about science as two as two two type of things you know there is something which is kind of an optimization and something which is discovery you always hear you know like uh human in the loop and whenever I hear human in the loop I'm not saying it's wrong but I'm saying human in the loop

00:38:28

 is not the right thing to do for for everything. So what is human in the loop first? Human in the loop is that you have models you know models that are working and and you know like you know proposing some direction but then you always have a human to say hey you will do this not that >> a bottleneck that make a decision. Well, the that's kind of the pessimistic way say a bottleneck. Maybe the thing is the more positive way is like well we we want human to decide on every every important stuff. But then I always I

00:38:58

 always argue that in some cases you don't really want human to touch it because you know it's very well known that a lot of the experiments in life science are not reproducible. You know I think it's it's been published everywhere. I think there was a a famous amgen paper from several decades ago where they looked at the oncology paper and I think 60% of the experiments were not reproducible. Then a few years ago nature also did a survey of life scientists in academia in the industry

00:39:26

 and said something similar 60% or 70% of of the experimental principle. >> So imagine if you have an optimization experiment. So what do I mean by an optimization experiment? Let's take an example. you have a a cell line and you want to find the ideal condition to grow the cell. So what does it mean is that you you need to find the right medium. Uh the right medium will be a mix of you know bunch of salt and liquid. Uh so maybe you have like four parameters like the con different concentration of the

00:39:57

 different salt different pH different temperature. So if you want to do that uh well then you need to create what is the experimental universe where you have to test things and where you do it you need to do it in a in the optimal mathematical way. So there is an optimal mathematical way. You can't test every single every single experiment because there could be hundreds of thousands or millions of experiment. But what you want to do is make sure that you test enough experiment that you could sample the experimental space in

00:40:28

 the the ideal way and computers are much better than us you know to to uh to do things to sample things in the right way. when I was doing my PhD uh in many ways when you do c uh you know when you try to create crystals because that's what you need to do to uh to be able to solve a crystal. Sure. The first thing is to grow a crystal. Growing a crystal is something that takes years because you know like finding the right conditions. You you have to sample thousands and thousands of conditions.

00:40:57

 And I remember when I was doing it, you know, it was really a brute force approach. I mean you you have a you have a robot. So initially we were doing things manually but then luckily we got a robot. We were trying experiment after experiment after experiment and we were not really thinking too much like did I use the right combination of salt at the right temperature at the right condition. We're just kind of >> using all the variables that you >> Yeah. going through all the kit we could

00:41:23

 buy and uh >> there there was way more intelligent method to to follow and to be honest this is the type of thing where it can and should be auto because I think this is really the type of thing where you don't need a lot of creativity. It's really following you know like a metabical approach. Now it's a different story when the very question you want to ask is unclear. You know like you want to do something you want to research a certain area but even expressing the question is is difficult. So that's the

00:41:55

 difference between optimization and complete discovery and uh I think the the complete discovery is still where we are there are examples where AI start to be actually really good at discovery and I can talk about that later on but I think it's still where you know the the human uh the human involvement is is very important but um you know I I kind of diverting because we were talking about why >> silicon medicine why are they >> why they are so good I think they've they've been able to use AI at so many

00:42:28

 stages of drug discovery and development. You know when people talk about an AI developed drugs, it's the definition is very is very [clears throat] clear. But my my own definition of what is an AI developed drug would be a drug that comes out of a program where AI was involved in significant part of the value added of of the discovery you know from I don't finding more information about the the disease biology, identifying the target of targets, you know, being invol involved in the design of a drug, being

00:43:03

 involved in the prediction of the efficacy and toxicity of the drugs, being involved in the you know, in the design of the clinical phase and uh in silico being uh you know like a tech bio, they are still relatively young. They managed already to move a number of assets to clinical phases. sure that either this year or maybe next year or the year later they they will manage to have um approved drugs because there there is there is no real AI developed drugs that that is marketed at the moment. So if you if you look into the

00:43:37

 literature you will see that there are maybe one or two drugs that are called AI developed drugs that are AI that that are on the market but if you if you you scratch the surface and you realize these are repurposed drugs. So a repurposed drug is a drug that was used against a different disease that you know [clears throat] developed against a different disease that has been repurposed or another. >> So AI can be used for that. But then for me it's a bit difficult to say is this really an AI developed drugs. No AI was

00:44:06

 involved. Yeah. But with in medicine I think it's different because they have a number of drugs that really like fit the the definition that I gave what is an AI develop drugs and where they really did a lot of work and where in medicine is really is really interesting and quite unique is that a lot of the stages that take several years they were able to compress it to months. So if you look at I think now on average they can go through discovery and pre preclinical stages in 18 months when normally take

00:44:37

 five to six years. So they've been able to demonstrate that they are able to do that. Now again I mean people who are uh let's call that AIPIC they will be they will say well but they still haven't brought any drugs to the market and and to be fair that's true you know that's true they haven't bring they haven't brought any drugs to the market. Now what I will say is well if you look at you know the traditional uh drug discovery and development pipeline is that all the company will take five or

00:45:05

 six years to go through a similar stage and since we have 95% failure >> there's still no guarantee that they're going to bring anything >> they will fail but they will fail after five or six years instead of failing after 18 months I'd rather fail after 18 months. >> See I want to highlight that because everybody it's a slogan in the drug development world. Oh, you need to fail fast. But like nobody's doing anything with the tools that we have to actually fail fast like what what you are uh this

00:45:34

 example you're giving um you know and if they have something approved amazing uh but if they don't they have managed to fail so much faster and go through so much more of this process with AI tools right this this failing fast was is there forever and it's kind of like oh let's do it faster faster, better to fail fast even if you don't succeed. But it's kind of like a slogan that doesn't really resonate with people because they hold on to this hope that oh if it's so

00:46:06

 far in the pipeline maybe I mean there is a higher probability that it's going to get approved but there's no guarantee right >> yeah no no exactly so I think if we are uh honest about where we stand today with uh with AI in drug discovery and development we've had a number of successes in accelerating some phases but if the ultimate success is actually bringing drug faster to patient we haven't done that yet I'm still very hopeful and I think we have a lot of strong positive signal that it will

00:46:40

 happen but you know one thing though where I will be like I'm I'm still surprised that we don't see more AI developed drugs in clinical phases so there was a paper in from the Boston consulting group that you know tried to to to look systematically at how many AI developed drugs were in clinical phases and it was I think 67 drugs in the end of 2023. I worked uh with an AI researcher Ed Greenblat and we actually were writing a paper on on that at the moment which I hope we will be able to publish soon uh

00:47:16

 where we uh so I asked him the question and then you know like uh he used all the tool that he had developed and the methodology to actually take the initial data set from uh from this paper and what happened 28 months later and I was really thinking okay you know with all the investment that we had in uh AI for drug discovery and development first like 300 400 drugs, we found only six or seven more. >> I was like, >> how come you know I was really really surprised and then you know like there

00:47:45

 are there are lots of there are lots of potential explanation why if you look at the initial list of 67 drugs you know some some actually failed. uh a lot of them change ownership because a lot of the tech buyers you know they they they will take the drugs only so far and then you know when the assets start to be interesting then it moves to a more traditional farmer that has the scale >> to take it to the next stages because >> yeah that's point >> yeah I'm not saying that large farmer

00:48:13

 don't have a yeah sorry just to finish on this thing I think I'm not saying that large farmer don't have a place I think you know that deix you know the CEO of Eli was saying large farmer very small. >> My goodness, Eli Liy is partnering with so many people. I get like alerts. Eli Liy did this. Eli Liy did that in AI space and this is what like let me let me let you finish your thought and then I pick up my thoughts. >> Yeah. Yeah. We we'll go back to that. But what he was saying in a in a very

00:48:42

 interesting podcast he was saying look there is something there is a place where scale plays a role that you know you can't you can't replace with anything. you know in clinical trial scale is still very important because you need to manage like a very complex supply chain problem and this says you know for commercial you know like when you're a multin multinational you know with sales force you know everywhere in the world then that's something that you can't really copy if you if you're a

00:49:09

 small company so however he also said that scale doesn't really work in in R&D you know like there is no positive there is actually a negative correlation between the bigger your R&D site in the world the least efficient and that's why he started a lot of partnership with uh with tech bios and AI companies and he's also you know trying to uh to bring AI within LI and of course I mean he's doing the smart thing now they're their stock price is so high they have so much

00:49:40

 cash they they are so successful that you know it give them uh you know the the mean to invest in a lot of things and to experiment in in any direction and to you know to make them more future proof because you you never know how long you know how long you will be successful and how you will be able to generate so much >> ride the wave keep riding the wave. >> Exactly. So they're really riding the waves and I to be honest I think they are doing a lot of writing and uh he's a

00:50:09

 very very interesting guy to follow because I really like his view on the industry and he can take like a very complex uh problems and and making it understandable for everyone because you know like you and I we know we know relatively well a corner of this industry. I mean he's a CEO so he has to know you know like >> knows the whole thing >> good enough understanding of end to end >> and this it shows a pattern because I'm seeing that in other pharma companies with acquisitions and we're going to

00:50:40

 talk about the specific acquisition in the digital pathology space but like what it it kind of follows the pattern that you just outlined the top of the funnel is where there's a lot more AI you have tech bio that thinks differently bottom of the funnel is regulated less AI more scale needed to execute and the pattern that I'm seeing okay those people who are those pharma companies that are successful at the bottom of the funnel which you know have multiple pipelines multiple drugs approved they are now partnering with

00:51:15

 companies that think differently with the tech bio companies or like trying to bring it in doing some kind of acquisitions which means like it's it's good leverage I see because trying to restructure like you said a company that is like hundred years old the the first u company that's like family led that comes to my mind is Binger Ingleheim this is they are so traditional h right and for a good reason and it served them well and it keeps serving them well in a specific uh in this this specific part of the funnel

00:51:51

 right but to capitalize on what's happening outside of their organiz organization these like external satellite partnerships where it's this where there's this synergy okay you can um run a lot of R&D on a low budget and only uh actually run wet experiments that cost more money that are super specific because you already modeled and predicted so many things and then okay once it works the giant can pick it up and synergize and then like take it towards the end of the funnel and like

00:52:22

 you said there is a lag uh what did you say seven drugs that were like AI powered that were approved not really >> not approved but that move to clinical phase so it's not even >> to clinical okay >> exactly so we don't see that but also there I always say there's no guarantee that you're going to find this biology right there is no guarantee that there's like more ways to cure these diseases the lack of guarantee should not preclude you from trying like one thing

00:52:51

 that is I like a lot in the digital pathology ology space are the computational pathology biomarkers. Um so I'm referring to the uh trop 2 biomarker where you don't quantify the IHC. It's a non small cell lung cancer and the test relies on a normalized membrane ratio of the IHC marker uh to what's in the cytoplasm. So like a pathologist looks at it, sees the cells, but no way will anybody calculate visually a ratio which I love because you hit the hard limit of visual physiology like everything that was

00:53:33

 before PDL1 her 2 all the other like biomarkers IC based my biomarkers where basically the question was oh is there a lot in the right space is there a lot in the tumor or not was a visual gueststimation by a pathologist. ologist. Now we reached a little like it's a simple calculation right but it cannot be done it cannot be guesstimated visually which to me as a believer in digital pathology means yes we found something where we need to digitize that ties back into you said okay h people didn't think the data should be reusable

00:54:12

 and now slide hisystopathology slide is a piece of data that everybody wants to be reused usable because you can do a lot with the image now with computer vision with AI powered computer vision. So anyway, but long story short, the uh the the collaborations of Tech Bio with the classical pharma, I like that a lot and I think that like you said, smart people are overseeing this in the giants and are on the lookout how to leverage things without having to test them themselves. >> Yeah. And and also to be fair I think a

00:54:48

 lot of it a lot of a lot of this is happening also within pharma within large pharma >> but it it just takes more more time. So I think it's it's also hedging your bets you know like transforming from the inside but it will take more time while working with companies that already have a working model and you know from which you can learn you know I was at a panel discussion with with a guy from Raj recently he was telling telling me that in his group they are making 100 millions prediction how I think it was

00:55:19

 per month. It was crazy but I mean you know like the number looks gigantic and it is a big number but you know like if you do like virtual screening of course you would have millions of molecules. So you kind of come up with lots of big numbers pretty quickly >> but um but then you know it means that when you do so many prediction it's not it's not doing PC's anymore. I mean you have that working in uh you know in in the in production. By production I mean really production in the sense of uh

00:55:47

 like a digital company will will talk about production. It means it's part of your standout workflow. Uh likewise I was at at a meeting uh you know it was another public academic meeting and someone from Astroenica was explaining how often they had hundreds of predictive models that they were using in the company for a number of stuff and they were explaining how often they they are updating those models and a lot were updated on a daily basis which shows also the maturity level of of those

00:56:18

 companies because in the past you know like updating things digital tools on a daily basis is what uh it's what Google will do or Amazon. So it's really like the way a tech company will think about that. So it means that in farmer companies now you have like entire part of pharma company that are working like a technology company >> you know that have people that could have decided oh I'm going to work for Amazon Google you know whatever but no I decided to work for a former company

00:56:45

 because I think the mission might be more interesting than you know creating yet the the next you know website to to buy stuff for social network I'd rather create drugs and then use technology to create drugs. Yeah. And see and it's u maybe it explains a little bit why why there was not really anything on the um and correct me if I'm wrong if you know about something but there has been I've been hearing at conferences from Google representative that there is so much happening on the Google health site. I

00:57:20

 don't see anything in the press but like you just said I see a lot of these uh tech bio acquisitions by pharma or some kind of pivot. So, so I want to because you mentioned Rash, I want to uh mention the recent acquisition. Uh Raj acquired Path AI and it was a 1 billion acquisition. Uh and we have this big pharma and we have this start pathology startup that part of their services was actually serving pharma. So this is something that is not so native for pathology startups because they are more

00:58:00

 like diagnostically inclined and making the life of diagnostician easier. this particular um company and I've been like following this company pathi for for several years like they were working with pharma so that means okay they know workflows they know pain points they know the direction and they found a partner that did enough meaningful work for them to lead to this huge acquisition obviously not all the acquisitions are that successful uh if somebody was following pagei acquisition also you know big milestones achieved by

00:58:36

 a startup. The acquisition wasn't that successful, but yeah, this is AI digital pathology drug development, which is pretty interesting for my digital pathology trailblazers, but do you have any opinions on this deal? Like why now? And what does this mean that this was uh this price tag particular? And I do want to contrast it with uh the page acquisition and feel free to give me your opinion on that because I don't know why. So first of all I I just want to before before we talk about that I

00:59:09

 just want to correct one thing on the Google health thing because okay if you look at Google you know it's part of alphabet and uh you know the example that I was giving with alpha before deep mind >> you know Google deep mine created a spin-off which is called isomorphics lab >> isomorphic labs belong to uh to alphabet and they just got uh another round of investment of 2.4 4 billion OVR. It's a massive. >> Oh, then this 1 billion deal is nothing. >> No, >> sorry. I under [laughter] asked. How

00:59:42

 could I have underestimated Google? I'm sorry. They >> Yes. To give Google to give Google the credit they deserve. I mean, they they are I think that's actually, you know, as much as I like in medicine and I think it's incredible and they have really great pipeline. >> Um, deep mind and now isomorphics. I mean you know they have guy two guys who won Nobel prize in chemistry you know as part of their staff developing algorithms they are one of the only the Nobel prize the Nobel prize let's let me

01:00:13

 give you a push back on this because I was working at Definians definians had a Nobel Prize winner as a founder they didn't do as well as the Google companies so Nobel Prize >> well they were acquired by Astroenica so it kind of fits the pattern but they didn't do as well as Google and they were founded at the same time as Google. >> No, but my my point here is that they are they have a staff which uh is quite unique in their insight and uh and uh >> you know like the the brain power around

01:00:45

 developing unique models AI models for for life science. I don't think that there is a lot of other tech bio that actually do fundamental research in developing AI models. A lot of them you know they will they will use model that have been public that are open source and combine them with very clever ways but I think isomorphic lab I mean they they do things like unlike any other company I mean that being said you know they are still quite quite new they don't have drugs on the market but this

01:01:15

 this is one of the company that I'm following quite closely because I I think the weight is different enough that it deserve you know speed but moving back to the to AI >> the acquisition in pathology and then rush. So first of all I think you know the interesting thing with AI in pathology is that uh one of the first modern application of AI not not in pathology or anything but you know it was image recognition you know like the the some of the guys that got the the Turing price like which is kind of no

01:01:50

 one in in AI uh in the AI world people who were uh some of the initial uh staff research staff at at Open AI I you know they they worked on the algorithm for image recognition >> and and in the end you know in what is pathology it's >> recognition pattern recognition >> ex exactly using AI to recognize pattern and um I think that's one of the one of the area where AI is the most mature at the moment is in pattern recognition so AI yeah pathology was you know one of the best area to uh to apply AI uh in in

01:02:29

 our field. So that's that's the first thing. >> It's an image based science image based medical specialty. >> No, I think the thing that is quite unique with with Rush, there are a number of things that are unique with Rush. It's a great company. They've been doing reorganization for a couple of years to be more AI ready. You know, there are always a number of ranking of what is the most AI ready pharma company. Think about think about it what you want you know like ranking in the

01:02:59

 end you know people tend to look at them and >> but search engine optimization >> yeah exactly but rush you know has always been in the top of those ranking and I think the last ranking that I looked at rush was at the top so that's that's one thing I think they they really understood that they need to do a lot of things to change to make AI work for them and to make a large pharmaceutical company prepared and and and you know like uh ready to leverage AI. It's one thing on

01:03:28

 the gent side. There is a very prominent academic researcher who started to work with them years ago and now is part of gent rev and a rev pioneered the the lab in the loop approach. The lab that's what we describe you know it's like you start in silicico you know you produce a list of uh of result if you look in term of compound for instance you know you will make a list of compound then you will test them produce data get the data back into the algorithm adjust the algorithm make another prediction then you do that

01:04:03

 in you know so she's really one of the person who pioneered this approach and Raj is uh you know one of the most prominent Third thing so ranking lab in the loop third thing is that rush is not just pharma it's diagnostics as well you know they have those two division >> exactly >> they are very very big diagnostic company so I think this is where you know this acquisition makes a lot of sense you know rush has always been you know at the forefront of using data in power using AI they have those two

01:04:36

 divisions uh so we can leverage an acquisition like that in so many areas you they can leverage it in their direct division. They can leverage it in their in their for when I when I saw it, I was like, hey, kind kind of makes sense. I'm not sure exactly what they will do about it. Will they integrate >> we're going to be looking? >> Yeah. Will they integrate it as part of their offering in the in medtech or diagnostics? Maybe. Will they leverage this as part of what they do in farmer?

01:05:05

 But >> I hope to get them on the podcast and uh and ask this team like what is it going to be cuz I'm super >> I want to get so many people from from [laughter] the podcast for so many things. I mean they they do great stuff, you know, like they are not doing like some kind of panel or or a drone thing. >> Yeah. I mean that's what I said. I I had a guy that was responsible for MLOps. Great great guy. Uh super smart and funny at the same time. So you know like the kind of best best type of person you

01:05:35

 can have on a podcast. So yeah I mean they are great. >> Okay. So yeah AI fantastic but what can go wrong and and something already went wrong. So uh and we mentioned this like oh documentation you that's like the easiest way to use AI to create documentation and that's what one company did. um and they get an FDA warning letter about uh not really reading through what they produce. So, let me give you uh a little story from yesterday. So, I'm uh writing a couple of publications right now and obviously

01:06:17

 heavily AI powered. I try to maximize the use of these tools and I also force myself to read through it, right? And yesterday [clears throat] I discovered that Microsoft Word can read this stuff to it. It was like the fifth time that I had to go through this paper and I'm like, I cannot look at it anymore. Can somebody read it to me? And sure enough, Microsoft Word can read it to me. Well, Poor Lea did not use that feature and they got this FDA warning letter uh to specifically call out AI misuse.

01:06:50

 And this is a cosmetic lab. And you know, I feel bad to like highlight, but that's an example. It's not that, you know, it's it's it's something you're going to run against. And if you don't have hard guard lace guard rails, then you can end up uh like them that they used um Chad GPT for your QC documents and didn't verify it. Submit it to the FDA and basically FDA said, well, you produced it with whichever tool, you're responsible, so here's the letter and

01:07:22

 don't do it again. Well, what does it tell us about where the regulators are drawing the line? That's one thing and kind of straightforward like if you produced it, that's your fault. But second, what like what are the guard rails that we have to implement especially in the regulated environment, right? And the moment you mention the abbreviation FDA, you are either in or heading towards regulated environment. >> But I I think you said it, you know, like uh whatever you submit to the FDA,

01:07:57

 I mean, you need to make sure that that you read it twice or three times, you know, like >> but it's kind of a general principle >> right whatever you put on social media. Oh, there was another story. I don't know. I don't remember who but uh somebody before a conference from a pretty respected company decided that they're going to amplify their presence at the conference and published an AI generated graphic and the comments like oh looks AI generated this is wrong that

01:08:27

 is wrong this company is wrong is like they didn't read like read and verify this is like so simple and yet at this high level it's not happening >> is it sad is it normal what do you Yeah. No, I mean to be fair, so I you know, I won't throw a stone at people and then not talk about stupid stuff I've done, you know, like in the past. I mean, >> oh, I did my share of stupid stuff as well. >> I've done something similar in the past, but again, I mean, the stakes were quite

01:08:57

 low because it was it was a a LinkedIn post. uh and in the LinkedIn post you know like I I took you know there was a table from a table from few years ago and then a table for similar theme but uh few years later and I asked an AI model hey tell me you know comparables to and then tell me you know which company went up which company went down he gave me a great graphic I posted it and I was like hey look I mean it's it's quite cool and then someone was like there are a few strange thing in your

01:09:25

 thing you know like this company wasn't there and then I started to look at it and I was like, "Holy shit." I mean, it's all made up, but it was so well made up that I, you know, I I was like I, you know, I did it quick quick and dirty, wrote a quick story and then proceeded without >> check it checking and then I was like, well, you know, that's my fault. I mean, I'm 100% responsible for not checking that. Now, the difference that is not something that I sent to a regulator.

01:09:55

 >> That's true. You did not send your LinkedIn post to the FDA. No, I did. Good for you. [laughter] Um but but but that's the thing you know like uh I I still should have done that because then you know you're professional you're doing something publicly >> you are 100% responsible for what you publish and where you put your name in whoever does it you know whether it's a human whether it's a colleague it's an intern it's a contractor or an AI and now I think the problem with AI is that

01:10:24

 it tends to do thing extremely convincingly so the other day I I really the same thing, but this time, you know, it didn't do it publicly. I, you know, I I used the eye to to uh create a list. And then I looked at it quickly and I was like, "Oh, it looks great." And I sent it to a colleague and I say, "Hey, look at the list. I mean, it's really nice. I mean, it was able to pick up new stuff." And then and then, you know, I sent it to him by email and then I I started to meet more of the list and I'm

01:10:48

 like, I sent him an email. I said, "Just delete my email." But but it was so well done and so convincing and I think this is the the biggest problem with a lot of AI at the moment is that it it still hallucinate quite a lot in very convincing manner. So some you know like where where it accelerate your workflow it's in the production of information where it's uh slow your workflow is that you need to you know you need to spend more time reviewing things. So you know in the past you know you were creating so maybe

01:11:20

 you were you know you were obviously taking more time creating something and you knew more what were going and then then you know rereading yourself was faster now the creation is like really compressed in a few minutes but then the checking can take way longer because sometime you have things that are so credible that you can you can miss them and I think this is the biggest danger but going back to this company you know there was another example like couple of years ago when JP just started I don't

01:11:47

 know if remember this story. There was a a lawyer that was arguing a court case and all the what's the name you know that you use in the legal term it's um example of previous judgment there there is a legal name for that obviously not the lawyers >> he had preceded um all the presidents or the jurist prredence >> jurist prance >> was completely made up and the guy you know like and you know like I think the judge found out and the guy got in big trouble and for me it's very similar to

01:12:16

 this example with the FDA is that >> whatever you use and submit in a in a high stake environment you need to check it >> and I'm sorry you know I was kind of I know it's bad uh with this especially with in the end it's linked to human health >> so I mean whatever you do you need to be responsible and check I'm sorry >> I'm glad they did I'm glad they because they had their own like adventures with AI on their end But uh that that's their job. But also this problem and I had

01:12:50

 another guest on the podcast, a colleague of mine, a veterary pathologist uh Candice Chu, and she says, "Yeah, this madeup stuff, the references that are made up. I'm guilty of that as well. That was a 2023 problem, not a 2026." like everybody now is aware of the limitation that it can uh make up things and all these like problems that we're discussing right now is basically okay not checking what you produce with the tool that you produced it with so kind of simple and I was covering the national comprehensive

01:13:25

 cancer network conference on AI using AI and the main theme was okay guard rails built in into the tool and you know in the um in the consumer AI tools. You can already um upload like skills, ways of doing things. You can ban AI from like inventing stuff. You can ask it to get everything from the source. And you know, I have all that already in place for my um scientific writing project in Claude. And sure enough, I did go through all my references and two of them were incorrect. Not like majorly

01:14:08

 incorrect, but uh I did have to manually correct that. So, you know, non-negotiable. Either you do it, somebody else does it. >> Mhm. >> If you if you were to write without AI, would all your references be 100% correct? I don't think so. >> You don't think so? >> I mean, imagine >> what does that mean correct? because I would not invent them. How am I going to invent? >> I mean, a lot of the thing a lot of the thing that I get wrong at the moment is not necessarily the reference itself,

01:14:39

 but the link >> no link. Yeah, that that's a that's a common thing like >> and that tends to you know at the moment especially with code I mean code got excellent for I mean I'm in absolute love with code but uh when I when it gives me references in general the reference exist and is good but then if I tell him well at the end of his paper you will build the you know like the references yeah and he will write it and then put a link sometimes the link the link is more so you know I checked

01:15:08

 everything and then you know I have to correct that was my problem >> yeah I have to correct a few things and that that's my point, you know, if I was doing things manually, I'm sure I would get a couple of wrong >> stuff would slip in there. >> Yeah. And and I think that that the other thing also we have a tendency to judge to be more harsh with with AI. We want AI to be absolutely perfect. And what what I always argue with people saying that is like are you always perfect? Are your colleagues always

01:15:32

 perfect? No. I mean, you know, like >> I love that you can be so harsh with AI because then you like don't have to hedge. So I have I'm I'm using this. You can basically like instead of yelling on my kids, I yell on my at my AI that's not doing well. >> My daughter does the same. You know, like sometimes she's using I should not say that, but sometime, you know, she's doing she's using Chipity to help her with homework and then she's talking she's talking to it. You know, she

01:15:57

 doesn't type, she talks. >> Some sometimes she's yelling at you. She's like, "Chachi, you're stupid. This is not what I asked you." >> It's incredible that we start to talk about it like that. I know because I have this um software. It's called Whisper Flow that lets you basically dictate everything. I love it. It's like a game changer. Anyway, so it makes an analysis. What's the tone you're speaking? What are the like how many words per month? And also what's the

01:16:31

 most common word that you're using? Guess what's my most common word that I'm using? >> Is it a swear word? >> No. No. know >> I don't know >> is the word don't >> because it [laughter] like does something I'm like don't do this don't do that don't like repeat myself you were supposed to do it correct the first place why don't you go to Google Drive on your own why are you telling me to search for something manually anyway so

01:17:01

 don't is the main word that I'm using with my AI but yeah so now that we're talking about the tools that we're using. H I obviously like I'm like let me just pay for the subscription, right? $20 for Chad GPT and Claude now became $200 a month because I wanted more tokens and to write more and you know whichever tool and it felt so low low price entry. what and like me myself I see that okay not every tool is going to uh fulfill your every requirement and like help you build the workflow so oh

01:17:42

 which next tool so then I buy a subscription for another tool and then I run out of token so I upgrade this tool a subscription I'm like oh next month I'm going to downgrade I did not downgrade yet so let's talk about the price >> yeah and uh I mean you know AI and how much it costs to you know try it out and will it work or not. >> Yeah, it's the same because I so now I know I'm for instance I I have subscription to uh personal subscription and work subscription to clone and chat

01:18:17

 GPT. So both personal and work >> um I have for my business person personal I also have Gemini. Uh I don't I don't >> Gemini. Yes. >> Yeah. Or not because it's No, I mean I use the free version. Um >> for some reason I I never >> Yeah, for Gemini. Uh it's also because it's it's linked to Google Drive. So now I find >> But you have to pay for one. Yeah. Whatever terabyte you pay in one way or another you pay. >> Mhm. >> Then uh you're using uh Riverside for

01:18:48

 podcast. I'm using the strip >> side flow >> now. You know, now I got, you know, I I was looking for an excuse to buy a big camera. So now I have, you know, for [laughter] my podcast I have this big camera. It's great. >> It's great. But >> this is a great excuse. >> I'm really edit. >> Yeah. But now I have 4K and I realized, oh, if you want to use 4K with the script, you need to upgrade your your subscription. I'm like, >> oh, really?

01:19:15

 >> You must be kidding me. Yeah. They want 100 bucks per year. Now I'm like, seems sick. And uh and and that's the thing you know like so for CHP now I'm always on the verge of you know like dropping the subscription but there are a few things >> but it makes images makes pretty good images I have to say >> images are great and I and you using whisper flow for me the the voice to text is with whisper exactly from from so I use judgity for that and my wife really likes judgity so I'm like okay I

01:19:43

 still need the subscription >> I think I know >> yeah the thing that I'm worried about is the you know everyone is talking about uh the fact that the prices will significantly increase but the way I'm thinking about open AAI tropical all those company they are like digital drug dealers dinners you know they they gave you your your tools for very small price >> and you know like a 20 bucks and then you know like then you you're hooked because you do everything >> and then at some point you will be like

01:20:13

 you know what actually you've you've been paying 20 bucks but for us it cost us much more so Oh, price would increase like five times, 10 times. >> Yeah, that's what happened to me. I wasn't at 20 and then I had to like write a lot and I zero, not even blink of an eye to upgrade [clears throat] to 200. >> Yeah. Yeah. Because you get value. I mean, that's the thing, you know, you but then even if you're 200, you know, there are some, you know, forecast saying that it will it will go

01:20:42

 significantly higher. And I think this is where >> have you ever tried Deepick? >> Not yet. Deepips is a Chinese model >> and uh when I tried their last version, the V3, now they had the V4. I haven't tried the V4 yet, but the V3 I remember was fantastic. At that time, it was really the best model. They just released the V4 and they decreased the price by 75%. >> They are the only company I know that decreased the price. Why? Because um since it's a Chinese, >> what do they want from you?

01:21:12

 >> They want your data probably. [snorts] >> Oh, well, because they are the one they don't want your data. I mean you know we can >> everybody wants your data. That's true. >> We can talk a lot about the e about the ethics of of all the AI community. I think that's an entirely different discussion. >> Yeah. That's a philosophical discussion. >> But then I think also you know like a lot of people in including myself have been using AI for a lot of simple thing

01:21:36

 and you know not in the most efficient manner. Uh and I think as the prices will increase we'll have to be much better. >> What is it that we use it for? Which model are we using? Uh are we using it in an efficient manner? So for instance, >> let me tell you there's gonna be AI that's going to do this for you and you're gladly paid. >> Exactly. No, but I think there will be at some point like a layer of AI that decide, hey, if you want to rewrite an email, you don't need OP 4.8.

01:22:03

 >> Yeah, you don't need that. uh if you want to do >> there already is I'm listening to different podcasts and you know they they tell you the best is like a YouTube video about how to do everything for free with Claude or whatever like for little money with Claude and then a sponsor is this like AI orchestrator that lets you use 27 different models for all different purposes and decides for you what's the best and there is another subscription on that. >> Yeah. And but then you know I think

01:22:34

 there are people also who will say look at some point you know you might want to have something installed locally uh and run open source model if you're technical enough. >> I mean you know the the computers so yesterday I think there was an Nvidia uh conference in in Taiwan where they talk about new laptop chips that will be able to run a lot of things locally. So you know think and think things are evolving so fast. I think you know what we see as a limitation now might not be a limitation in in six months but if the

01:23:04

 price imagine the price increase by 10 times tomorrow then you know we need to be very creative we need to think about which model to use you know using smaller models using things more efficiently not using it for rewriting email not you know today for instance I'm uploading pictures and powerpoints and PDFs and I just put everything it's inefficient >> and and then you know like when I started is enclosed, you know, at 9:30 in the morning, I will run out of token and then I will have to send an email to

01:23:34

 an administrator say, "Hey, can you give your can you give me more tokens?" >> Uh, and and now I'm like, "Okay, I stop being stupid." You know, I stopped uploading 10 documents. I I'm not uploading a PDF, but a markdown document because I started to read what to do when you're running out of token after an hour. Like, you need to think about the type of of document you're reading. What are then you need to think like a computer scientist you know what are

01:24:01

 what are documents that are more computationally efficient >> all sort of things that to be honest I mean we were not trained to think about that but at some point you need to be to get educated because that's that's we >> need to learn right yeah >> yeah but that's quite cool >> yeah I know so any like what would you recommend at the drug development scale how to get involved D with AI and on the other side, okay, consumer of AI and we kind of started, okay, you need to be a

01:24:33

 little bit more savvy, but like where we sit right now when you run out of tokens uh at 900 a.m. already uh or where like the experimenting phase for free, what do we do? Do we get more AI? Do we get different AI? Uh one um accusation uh and people from the computer science world are like accusing of measuring the gain that they are not really return on investment but the like that like the more you use it the better right the the the goal um in some of these companies is oh if you're not using enough then

01:25:20

 you're not maximizing izing its potential. Use more to get rid of human labor. Any final thoughts on the new way of using AI and drug development and >> consumer? It's uh it's okay. Uh I mean it's it's such a it's such a broad and and complex topic, but I will say potentially yes, you know, like using more, but it's not just more, it's better, you know, it's it's being smarter on how you use it. It's again it's not just copying what you would do

01:25:51

 and just you know putting AI instead but it's it's thinking differently. It's the example I was giving you about this co who says you know like I need to unlearn a lot of things and then relearn how to do it in the in the agent world how what is it that can that I can outsource what is it where I don't need to be the one doing it deciding what is it that I can automate and let an AI decide potentially without me being involved all the time because I think there would be also one thing you know like a new

01:26:20

 disease like agentic burnout you know like jet jeten the CEO of Nvidia would say, you know, with with agents, uh, agent will harass you. They will, you know, ask you for feedback all the time. And, and I see that, you know, when I use >> Claude does that. It like tells you such a huge message for like one little question like, >> exactly rewrite the skills. >> I kind of I kind of like how Claude does, but sometimes it's almost too much. And I'm like, okay. You know, like

01:26:48

 when I talk to someone, when I have a dialogue, I don't I mean, maybe maybe I do. I don't lost you 10 minutes but so [laughter] >> we can analyze this podcast how many what what percentage of time no worries >> um but yeah I I mean I think things will evolve and they are collecting so much data that I'm pretty sure they they are well I'm sure that they are evolving the tool you know and they are they have in the history of digital tools the best understanding of user behavior because

01:27:19

 you know like the difference with Google in Google you know you used to search like one word or short sentence here you write your whole life in this thing. So yeah like the context they have is like unlike anything else. So the tools are evolving super fast but if you if you're talking about uh that in the in the context of organization of pharma companies I think I think there you can't just give access to your uh researcher and expect them to be to be efficient. I mean it's part of such a

01:27:50

 big transformation program you know you need to you need to uh so need to start small get you know early win scale train your people I mean there is a huge investment in upscaling people and getting you know getting technologist and data people together with with scientists you know telling scientists to change the way they they work have always been working is almost impossible you Some of them will be you know like into it enough that they they've done the work uh themselves but a lot of them

01:28:24

 will be skeptical and because we've been trained to be skeptical. So unless unless you you put in place the right change program in the organization you won't see any ROI because you will have resistance and then you will say oh you know like societies they are old school. It's not it's not their fault. I mean it's it's it's you're changing a complex organization where people always think about the guardrail you know like the fact that we work in regulated environment so it's not just you know

01:28:54

 using tool because it's food and everything you know it's working tool is very serious business where the consequences are high expensive we are developing things for human health so all those things are are important and scientists have all that in mind so changing the way they work means that you need to have comprehensive programs in place and bring them in a in a journey and I think that's also one of the reason why changing things in farmer is complex because it's a very complex

01:29:24

 environment but I think a lot of the pharma company they decided okay let's bring some of those company that have done it from the getgo you know the tech bios let's work with them and I think what will happen is what we've seen with biotech you know biotech got acquired by far we will see tech bio being acquired by by At the moment they have strategic collaboration eventually they would be acquired and then there would be you know like an excellent center of AI drug and development and several years from

01:29:53

 now you know things will will start to be you know like in the organization they will learn from that they will hire profiles of people that we normally don't have you know more technologist you know they they will they will have more people who you know as I said earlier in the podcast could have worked at Amazon Google whatever but our team so that's what I think would happen. >> Interesting. So do you have before we wrap up do you have the most annoying quirk of any of the AI tools that you're

01:30:26

 working with? I have one specific one that I >> What is the most annoying? >> Well I I think I think it affects you you mentioned the word don't. the word don't you know like I I I so I trained I trained Claude to uh you know on on my specific style of writing you know I went on the on the hunt of all the content that I have written before AI so I went my old emails yeah I took my PG you know that I wrote the my PG is the longest piece of thing that I wrote before AI >> and I uploaded all of this into pl and I

01:30:59

 said this is how I write and and then I said you know like I don't want to have too many m dash because I like that. I hate absolutely hate when you have a sentence that say uh this is not this but this you know like this >> oh my goodness I discovered that one I put it in my skills it always like I was writing a publication >> and then so you ask what what annoys me is that when despite all of those things it comes back in the writing and then then I tend to to really yell at it I

01:31:27

 say I told you know in in capital letter I told you not to do that and then he apologized and I'm like seriously Exactly. So I think this one and then you know the hallucination the hallucination even though it's going down you know once in a while you have something that that's crazy and that should not be there and that's getting more and more difficult to so you know the stylistic things is annoying the >> hallucination which is difficult to to spot I think is more serious and this is

01:31:57

 the one thing where I'm like >> that that really that really really annoys me because sometime you can quickly lose credibility if you have something which is you completely >> putting out something >> and the fact that it's becoming more and more difficult to spot. That's something that annoys >> Mhm. >> a lot. >> I have one additional thing. It's just mostly annoying. It doesn't keep track of time at all. It doesn't know what date it is, what time it is. Like I'm in

01:32:27

 a chat and I'm like, "Okay, today I'm going to do this and that." And it's like, "Oh, you've done so much today." And I'm like, "No, I've done this that you're mentioning yesterday or the day before." And it has like zero context of time. I don't know if there's some like button to make it time aware that it has an internal calendar, but I didn't figure out. >> Are you working with projects? Are you working with projects?

01:32:53

 >> Yes, projects as well. Because for for me it worked quite well with project where I kind of uh tried to wrap up big project in you know in so you know they they call it bible you know when you start to write a super super long thing and it call it always tell me I will make your bible I don't know why it call it like that and the what it called what it call the bible is kind of the the the document that kind of links all the thing we've done in the project and then the different part of the project that I

01:33:24

 tend to doing one day by one day are kind of linked to this thing. So that's kind of how it's not necessarily keeping time but it keeps the stages of a project uh quite well which is almost equivalent to time. >> See equivalent I want the time not almost equivalent to time. >> So it's it's always working around it's always working around the current limitation and quirks of I think that's that's what we learned about uh you know how to do. >> Yeah. Thank you so much. Thank you so

01:33:55

 much for joining me. It was so much fun talking to you. Uh I'm definitely going to include the papers we mentioned in the show notes. Also your LinkedIn profile. Pretty inspired how you explain research and have you write. So let's see if I can model some of that. Thank you so much. >> Well, thanks Alex. I think it was fun. It's a it's a long it's a long recording. So good luck for >> I know. And uh I need to also uh link to your podcast so that people can find you

01:34:24

 online. Uh tech and drugs. >> Cool. Thanks a lot. And uh yeah, I mean uh let's do it again in a couple of months when you have new stuff to talk about. >> Do it. Exactly. Thank you so much. Bye.