00:00:04: The Biorevolution podcast.

00:00:06: Your hosts?

00:00:07: Luise von Stecho

00:00:08: and Andreas Roichler, the biorevolutions podcast this time lab in the loop.

00:00:15: AI bridges to the wet lab.

00:00:18: we have a fantastic guest today at the table.

00:00:21: easy but as always We start with our quotes And then you do the kudos on the introduction of our guests.

00:00:28: I guess

00:00:28: exactly.

00:00:29: let's do that.

00:00:30: We have two quotes today.

00:00:31: One is from Vifre Gaff, from Genantek and she's like a pioneer in the systems biology field.

00:00:36: actually I was reading her papers when i myself were still in The Lab which is a long time.

00:00:41: so there's a long way to bridge to the wet lab.

00:00:44: for me it's almost Fifteen years by now,

00:00:47: so I'm twenty years.

00:00:48: Oh it's a

00:00:50: long time wow!

00:00:52: But what she says about the partnership that Genentech struck with NVIDIA some years ago AI The Lab and clinic together to uncover otherwise inaccessible patterns in the last quantities of data to design experiments, so that already gives us a hint on how that could work.

00:01:10: The other one is from Jeffrey Van Malsen, the flagship general partner and head of Lila Sciences – another lab-in-the-loop company.

00:01:18: I'll only read this quote at an endpoints article because it's very nice metaphor for building a beautiful super intelligent mind requires science and that I think refers to the love-and-the-loop approach, which also very much like.

00:01:39: Happy to have you here Thibault!

00:01:41: Thank You for having me.

00:01:42: yes

00:01:42: your podcaster yourself so we will learn i think some interesting things about guests on our podcast who are at the intersection of tech and Drugs, which is your newsletter in your podcast.

00:01:55: Which I think it's going very successfully And i think It's Very interesting to have a conversation with this fellow scientific communication professional who can give the insights into This super-interesting emergent field.

00:02:07: so you Have been In this area between life sciences technology and publishing also in The past for over twenty years.

00:02:16: You're now a silica group but you Also have this angle of Communication With A very active link in profile with the podcast and newsletters.

00:02:24: So happy to have you.

00:02:25: Thank you for having me.

00:02:26: I think it will be a very interesting conversation today, i hope It's not tech and drugs and rock'n roll but it is Tech & Drugs.

00:02:34: Well when I was designing the logo When asked my agent to design the logo You know?

00:02:39: I wanted one of these retro type things that we had in an American diner.

00:02:44: so... Oh yeah!

00:02:45: ...it´s a very rock-and-roll field.

00:02:47: Some people might not see this way But its super exciting feel at moment Absolutely Can explain why.

00:02:54: Why?

00:02:55: Because, you know I think okay so that could be an entire podcast on this topic.

00:02:59: but when people talk about having a mission.

00:03:04: When you work in something like...I cannot think of other industry where there is more important missions than life science and pharmaceutical industry because at the end we are developing drugs or patients And most of my talks say they have around twenty four thousand known diseases.

00:03:23: fourteen thousand are rare diseases.

00:03:25: We have between three and four or five-thousand drugs on the market, but those drugs are not for five thousand disease.

00:03:31: they're out for far less because you may be ten drugs against the same indication And then from both four to five thousand You maybe two to three hundred that actually curing the disease.

00:03:42: Every year there is fifty new drugs let go through regulatory approval.

00:03:47: So if we go at this pace We need another five hundred years until we cure all diseases.

00:03:53: Yeah, good statistics!

00:03:55: I would be before we dig into the AI field and were extremely interested in that.

00:04:01: but at the same time when you mentioned this –and i have to ask this question– where are the crossroads?

00:04:06: Where healthcare systems are on the edge?

00:04:11: The question, the basic questions is how on earth are we going to pay for this in the future?

00:04:17: And at same time you have this revolution that's going with new findings and new drug developments.

00:04:23: Quicker drug development probably under a billion euro per drug.

00:04:28: but those crossroads exist.

00:04:31: so we need to ask ourselves where do want be headed?

00:04:37: Yeah, I think there are multiple ways of looking at this problem.

00:04:39: So first off all i agree with you.

00:04:41: if we just continue to do the things that way it's not gonna work.

00:04:45: plus We're getting a lot of innovative treatment.

00:04:47: You know like yesterday was listening to the radio like normal Treatment as your aging?

00:04:52: There is.

00:04:54: five years ago That would cost eighty thousand euro a year and now I think since its in France from official number we have something like forty or fifty even more.

00:05:04: So don't quote me exactly on that, but that's really increased a lot.

00:05:08: Then if you look at cutting edge treatment like gene therapy how many treatments exist now?

00:05:13: That are over a million a year.

00:05:15: I think the most expensive is four million per year.

00:05:17: so... Incredible numbers!

00:05:20: Not sustainable But those things literally life-saving.

00:05:23: You know they have an injection to cure disease and kill someone.

00:05:27: It's true innovation The life saving thing for which i'm in this industry because it was wonderful.

00:05:33: However, to your point we already have huge deficit in all the healthcare system.

00:05:38: So I think a couple of things to consider.

00:05:41: first preventive medicine.

00:05:43: most of our systems are designed to treat instead.

00:05:50: that's the future of medicine.

00:05:54: You know, like do something before you get sick.

00:05:57: here I mean in Germany if you go to your doctor and say hey you know i'd like to do a blood panel every six months with lots of analysis can we do it?

00:06:05: They're like no!

00:06:06: If you want to do it you have two pain.

00:06:09: so for me its moving more preventive.

00:06:13: but then dig into a lot today.

00:06:16: It's if drugs, you know like the statistics for drug is typical.

00:06:20: statistic is ten to twelve years of development and ninety five percent failure rate The number that everyone using in terms cost two billion but actually most up-to date numbers are six billion.

00:06:29: when we look at top sixteen or seventeen pharma companies.

00:06:32: We need get this number way smaller.

00:06:34: I mean it needs cause maybe couple hundred million.

00:06:37: it needed takes three year not twelve years.

00:06:41: Then eventually it will be something much more sustainable and you'll be able to treat way more diseases.

00:06:47: For me, that's the two-direction where we need to go.

00:07:02: Super exciting!

00:07:03: Let's dive maybe a little bit into the impact that AI has on drug discovery development science itself?

00:07:10: It is interesting because this was our first episode so like... We've been talking about all of these things the AI things that we talked about.

00:07:21: And it turns out, I think in almost twenty episodes where an AI topic so this is really what's happening and bio tech right now?

00:07:31: It's also interesting because when back on listen to predictions like four years ago some of them

00:07:37: how good were you with your prediction?

00:07:39: Well!

00:07:41: The interesting thing was before large language models became a thing.

00:07:45: Back then natural languages Processing doesn't work that well.

00:07:49: And then suddenly it worked very well and things changed drastically.

00:07:53: because of that, we did not predict this but I think many other people didn't either.

00:07:58: But would you say like the impact that AI has on drug discovery development in industry are overestimating or underestimating it at a current stage?

00:08:08: That's a great question.

00:08:09: And I consider myself as an AI enthusiast, but also try to be an AI realist.

00:08:15: so i think its such a hard questions because you can look from many different angles.

00:08:20: So if we really tried being rational and looking the facts there are couple of companies that have demonstrated very strong signals showing What is it that AI can do to the discipline of drug discovery and development?

00:08:35: I write very often about in-city co-medicine, because in city co-Medicine has a really big pipeline.

00:08:41: And they've been able to demonstrate.

00:08:43: you take preclinical steps which normally takes five or six years... ...and compress them with eight months but on average eighteen months!

00:08:52: So it's, you know like less than half of the time to do it.

00:08:55: And also in terms of cost I mean they reduce the costs dramatically.

00:08:59: now if you want to be a bit more critical and If you zoom out You say well how many drugs actually that AI developed drugs?

00:09:06: We'll go we'll define what does it mean an AI develop drug?

00:09:09: They think It is very important To Define That.

00:09:11: But How Many AI Developed Drugs Made It The Market at the moment almost known.

00:09:15: So depending on how you look, so there is this paper from Boston Consulting Group that was published in the end of... Yeah we talked

00:09:21: about it a couple

00:09:22: times.

00:09:23: It's a great paper and actually I was mentioning another podcast later today where we took these papers and looked what happened twenty four months later?

00:09:32: So i'm working with scientists who are very interesting.

00:09:34: There really nice system were he can do like very interesting search and validation And an interesting thing is not much more came Like if you try to extend the list.

00:09:46: So I wrote an article recently, where are AI developed drugs?

00:09:51: That's one thing and what i was mentioning from paper.

00:09:54: just go back a little bit.

00:09:55: there is not lot of drug that made it market.

00:09:57: so in Boston Consulting Paper There Was One Drug Marked As Marketed But It Was Repurposed Drugs Not A Drug That Was Discover Where The Target Was Discovered.

00:10:09: So It Was Drug Repurposing Which Is Actually a great use case for AI, but it's not like from target discovery all the way to commercial.

00:10:17: I think there is another one from Lenten Pharma that made into market with data will have also two double checks of as one thing that we'll discuss isn't that?

00:10:26: Isn't that also due to the fact that AI hasn't been around simply long enough too?

00:10:32: make it commercial

00:10:34: say that certain applications or example in medicinal chemistry used older forms off for a while already, right?

00:10:41: So it depends really where as you say.

00:10:43: I mean that's the definition.

00:10:44: What is use of AI?

00:10:46: but- Yeah

00:10:46: its not just LLM.

00:10:48: Exactly!

00:10:48: But the generative AI and deep learning based AI we're using.

00:10:52: i think this hasn't been around for that long so unless suddenly time lines would be compressed super significantly We wouldn't even expect many AI discovered drugs to make into market.

00:11:07: Yes, you definitely have a point.

00:11:08: I think again if we split things in two i'm not surprised that all the embedding of AI tools into drug discovery and development process hasn't made it yet or haven't had an impact on large pharma because this varies a lot of inertia.

00:11:26: It's complex to change processes.

00:11:28: however why am saying i'm surprised is that We still have a lot tech bios And I would have expected with all the investment going into tech bio to see more of their pipeline, making it to a clinical phase.

00:11:41: Maybe again is there other optimistic me who wanted to see him or and i think we should not draw too many conclusions now?

00:11:49: The critical time will be what happens in the next five years.

00:11:52: because no It's really moving from experimenting with AI To bringing AI into production also in the large farm are like what will Eli Lilly do in the next five years?

00:12:02: What Will Roche Do In The Next Five Years, AstraZeneca as well.

00:12:05: Because these are really a company that now going full speed real investment and have like really teams using AI day in, day out.

00:12:15: So that would be interesting to see if we see any impact on their pipeline and the pipeline velocity in the next five years?

00:12:21: And I think the interesting thing is also that bigger pharma companies they're not as public about it.

00:12:25: for the tech bios its there language right.

00:12:29: They need to put a little step outside.

00:12:31: but for big pharma they might just do alot of things with AI and don't like communicate AS much A press release about every single step of the way.

00:12:43: You're right, not as much but in conference they do because I think there is now a lot of discussion also About are The large farmer getting the right talent?

00:12:51: So i was at a panel discussion At the swiss biotech day a few weeks ago and it Was one Of my panelists with an mlops lead at roche who's saying that They were making several hundreds of millions of predictions per week In his team.

00:13:05: so it really like proper production level AI, and now yes we need to have the right talent because it's not you know I'm a biologist by background.

00:13:14: You know?

00:13:14: It's not me that they need...you know They need someone who understands computer science Who yeah has a feel also for biology.

00:13:21: so That's why i think The role of translator are people who can understand both domain is becoming more important And Pharma is still not necessarily really, really good at getting those people or making the pharma like they're one place where people want to go.

00:13:46: Which brings us to the next topic.

00:13:47: maybe do we still need human scientists?

00:13:52: I was just having this exact same thought that in the day of AI and AI co-scientists.

00:14:00: what the heck are you talking about talent anymore right?

00:14:05: It's a great question.

00:14:07: If you ask me for today and in the next five years, I still think that we need more talent.

00:14:14: AI companies, you know whether they are in life science or not.

00:14:19: They still hire quite a bit.

00:14:21: what happens is that we're changing their talent mix.

00:14:23: so there's A lot of letting people go but to steal.

00:14:26: You know like a lot of company who are letting a lot Of People Go?

00:14:29: We also hiring large number of people.

00:14:31: So I think There Is a Bit of a renewal of the type of people they need.

00:14:36: But now if you ask do we Still Need people?

00:14:38: i Think you need To consider two aspect at least.

00:14:41: today varies on one hand optimization and then On The other hand Discovery.

00:14:46: For optimization, I'll argue very often that we don't want human in the loop.

00:14:51: If you take something like okay... ...I want to find an optimal condition for a certain cell line.

00:14:58: So what is it?

00:14:59: You have four or five parameters The temperature of pH concentration of different salts And then you need to change those parameters and see what happens.

00:15:09: Do this in a loop.

00:15:10: To be honest most people if they intervene will mess up.

00:15:16: Yeah, I mean because we know that... We have this problem of reproducibility in life science and you know i've seen it my own PhD.

00:15:24: You know like one of the key experiments during my PhD?

00:15:27: I don't if should say publicly but was almost impossible to reproduce.

00:15:31: It's such a headache.

00:15:32: And then when realized actually with some contamination in the buffer which came directly from the factory it took us months to realize That once

00:15:41: had algae growing It was sealed, and you open it.

00:15:46: And then suddenly there's this little fluorescent algae in the...

00:15:50: You realized that they were essential for your experiment?

00:15:52: Yes!

00:15:53: Well I wasn't studying cells anymore but studying algae at that point.

00:15:57: But yeah, it's interesting.

00:15:58: And I think one interesting conclusion of what you say is that a lot of the tech bios or also like this for example arc Institute are investing into producing their own data.

00:16:09: now right because It's taking just biomedical data from public repositories and From The Literature.

00:16:18: let robotic system do data production under really controlled conditions rather than trying to fit all this messy data together.

00:16:25: And it's not just that, I mean a lot of the public data is positive data and not negative data.

00:16:30: You know i worked for publishing company where was on the data side in development or digital tools but alot of uh...data are publication like everything actually works.

00:16:42: so have an eleven year old daughter And when she's complaining about things not working, I say look you know in my job we fail ninety five percent of the time.

00:16:49: So that kind of normal.

00:16:50: but then this ninety-five percent is invisible because that's not what people want to read.

00:16:55: however That's where algorithm wants to read.

00:16:58: You know the algorithm won't those ninety-fifth person Not just a fifth person Because you need to train them with What works and what doesn't work Had a project actually With FDA In one Of My former company Where they reach out To us They said Look We Want to Train Models To predict a pathotoxicity and the problem as we have FDA that no one send us thing.

00:17:17: That's trigger apatotoxicity because then, We will tell them without what you're doing it is quite bad.

00:17:22: or they had all of positive signal but they say can You share with us?

00:17:25: The negative signal?

00:17:26: And off course for my company has plenty of things that you know trigger apato toxicity.

00:17:31: But when you enter also into all the discussion on Well, do pharma company want to share their negative data?

00:17:38: And I think this is one of the things which are changing at the moment.

00:17:41: There's a bit more willingness to share.

00:17:43: you've seen that with Eli Lilly and their tune lab where they're sharing a lot of data.

00:17:47: i think there was a partnership between BMS and Takeda.

00:17:51: we were doing alot of data sharing so lots more pre-competitive stuff.

00:17:55: but just go back on your question optimization where I think you don't need people anymore.

00:18:02: You do not necessarily need human, but then there is the discovery part and the discoveries were at that moment.

00:18:09: i'm not saying what will happen.

00:18:10: in six months we'll discuss about AI co-scientists who are already doing interesting things in this area.

00:18:15: But it's still important to have people asking those fuzzy questions because sometimes It's even difficult to formulate the question.

00:18:25: And if you can't formulate a question properly, then it is difficult to give the right context for any AI system.

00:18:30: so there are still a bit of back and forth where we need to experiment and leave space also for serendipity or accidental discovery.

00:18:40: So I don't know whether an AI system could discover penicillin.

00:18:45: This type thing might happen.

00:19:06: How would you, because I mean this term AI scientist, AI co-scientist whatever it is flying around a lot.

00:19:13: And i think there's different things that different people mean by its.

00:19:17: and we talked about last year when this Stanford virtual lab then Google co scientists came out almost exactly here ago.

00:19:24: they were talking about but how do define what does an AI scientist?

00:19:30: Okay, so I'll give you an example of a discussion i had recently.

00:19:33: So there is this company called Edison Scientific that released the product called Cosmos recently.

00:19:39: And when they released it, we also release the paper where he did a partnership with wet lab scientists who took their AI co-scientist and tested it in the lab.

00:19:47: So I had an interview soon to be released.

00:19:49: When i'm done with editing or my partner is done with Editing Where I asked him you know what Is It that He Did With That?

00:20:00: So we had a dataset which was the gene expression data set.

00:20:06: It took this gene expression dataset, fed it to Cosmos over AI coscientists and then Cosmos started to do analysis... ...started to generate ideas like say oh I see that there is don regulation of genes.

00:20:21: so my first idea is another cell type must be upregulated check the idea and realize no, actually both are done regulated.

00:20:30: So it went and fetched other ideas from their literature an other data set And then started to generate ideas test them Generate more ideas, test them.

00:20:40: The way scientists were describing this is that It's starting like an octopus but with thousand arms and testing so many things in parallel at the same time To a point where he became almost scary.

00:20:51: because although you know This type of system will describe how I'm going do these?

00:20:57: It does so much and in parallel that it becomes almost scary because you can't follow.

00:21:02: You know, comes to a point where human brain cannot follow And I think this is really the power of AI co-scientists This ability of testing and connecting things That a human brain could not do.

00:21:15: It's exponential.

00:21:16: Then building new theory And then I think you can take it to the point where for me, The most interesting and we'll talk about later is if You can connect that to a lab and generate some data To test them.

00:21:29: because at the moment.

00:21:29: The bottleneck Is That It will Generate an idea?

00:21:33: When it's running out of papers or data set it Will tell you well now this is your experiment But i'm proposing

00:21:38: your turn when

00:21:39: you need to generate.

00:21:40: Yeah, you're kind Of serving the robot.

00:21:42: to finish the narrative here as he said After, you know like all the analysis it came up with a new theory and said okay my theories that this cell type of the cell population in Alzheimer is dying because there's an enzyme which will control the membrane flipping mechanism that exposes protein at surface.

00:22:02: Which is recognized by a cell type to kill these specific populations.

00:22:06: It was a completely new biological mechanism, which wasn't known.

00:22:10: Again it's not something that will give him an Nobel Prize but the type of paper people publish contribute to improving knowledge little by little.

00:22:20: then they took that and went into lab tested.

00:22:23: after couple months realized this is absolutely true.

00:22:26: so for me thats definition of AI co-scientists.

00:22:28: thats really the type thing that can take data set generate idea test them in a certain way and then, you know do this in an iterative manner.

00:22:38: And they eventually connect to the lab even be able to test things on its own or partner with human.

00:22:45: that's next thing we need to do.

00:22:47: what researcher was saying is it has completely changed his ways of doing research because now he's more managing a bunch of agents than... In the past for instance will have taken his gene expression data set spend hours or days, you know doing the analysis himself.

00:23:04: So it's really like compressing the whole research workflow in ours rather than days and weeks.

00:23:10: so at the end of day its iterations on steroids yeah.

00:23:14: And I think something also which is important not just reiteration but connecting things that are so far apart That normally will have a hard time connecting yourself.

00:23:41: We had an older episode that we called Expanding the Option Space, I think.

00:23:44: That's exactly what it is doing – giving within The Rural Book more options!

00:23:49: I have a slightly philosophical question on this.

00:23:52: You started with these nice statistics of how many diseases there are and how much drugs they are….

00:23:57: What was your connection between them?

00:24:02: There were two ways to use this idea pattern driven drug discovery.

00:24:08: using AI, using automation could go.

00:24:10: I mean on the one hand you could imagine that your follow down avenues that humans didn't think of and you discover truly new things.

00:24:18: in other hand we just had this example a narrowing of literature space basically and I think this is something that was interesting study out in nature.

00:24:31: The other day when they said the scientists to use AI become more productive but overall, over time to like simplification, to a smaller denominator.

00:24:59: Would you say it's going in the broader direction?

00:25:01: So broader option space?

00:25:02: or would you say its growing in the narrower direction?

00:25:05: Or we just don't know?

00:25:07: I think...I think We Don't Know.

00:25:08: To Be Honest!

00:25:09: Yeah i see where It comes from because also if You look at language A lot of people are saying that using LLM to write tends to unify and harmonize how People talk.

00:25:21: I think that's definitely one of the challenge, but is it a challenge because people are not using the tool in the right way?

00:25:28: Or Is It Inherent to The Tool itself.

00:25:31: One other thing i've done recently... Because Of course I use AI for so many things For Writing That I Went Back To you know, the pre-GPT time to find all of my writing.

00:25:42: You know that I had in my emails...I took my PhD dissertation and i loaded all that into a cloud to say can you extract my pattern?

00:25:51: And my style?

00:25:52: so..you know ..i think it's also matter how this is trained.

00:25:55: then can you train it another way ?

00:25:57: So that its deviating from the mean again ...i'm thinking we'll have two years to ask computer scientists.

00:26:04: maybe it's just also a phase, you know in those tools.

00:26:06: I mean we are aware of that and we're aware of the limitation And i think these things require more research and development.

00:26:16: Again like things evolving so fast The thing was a limitation month ago is not a limitation anymore today.

00:26:23: So im quite optimistic because people are aware about this and there is so much investment going into such tool.

00:26:31: you know, like a lot of the things that we considered barrier or limitation are being overcome quite quickly.

00:26:37: Aren't people going to say rather sooner than later?

00:26:41: Well I'm fine with that even if it's a limitation... I mean people who use LLMs at certain points they're okay.

00:26:50: i make my peace here.

00:26:51: and why bother?

00:26:53: this

00:26:53: is like in the first episode We had on chat gpt just after Like it came out, we discussed with a mathematician about and what the predictions.

00:27:01: And I think there are certain text-heavy tasks that are slightly nonsensical or maybe not the right way to make sense of world like given ourselves just read kind of the stuff that nobody really likes to write and read.

00:27:19: And maybe this is what the LLMs will take over, and it's also fine.

00:27:23: I mean there are certain areas like for example writing essays at university?

00:27:28: This just something we need to think about how to integrate AI in a smart way because It's little bit saying yeah no i don't want use the internet.

00:27:36: There was no way back.

00:27:37: now We have to learn how deal with it But at same time does not dumb down everything, but that we actually use it in a way to expand our intelligence rather than make us all sound like JetGPT or Claude.

00:27:57: But I think we are sitting this learning phase where you know they're... It's not the straight line.

00:28:03: and then at some point people realize okay.. That works!

00:28:07: That doesn't

00:28:07: work!!

00:28:07: You know?

00:28:08: That makes us more stupid.

00:28:13: You need to experiment, see what works.

00:28:15: What doesn't work?

00:28:16: The difference between this technological revolution and the previous one is that it's going so fast.

00:28:22: I think people just have a hard time coping with space of development.

00:28:26: but when you were talking... ...I was thinking about something that i've seen recently Talking about the tasks that people don't like Work email.

00:28:34: And then there are these colleagues who send very long work emails That we don't want them read.

00:28:39: So I had a colleague recently who sent with his long work email, he say I'm sending you this Long Work Email but there is a markdown document that can download so making it easy for you to put into your AI tool because people will not necessarily read.

00:28:56: So the thing which's interesting now is that the AI chatbot whatever one are using has become... The things we use to produce content and also to read content.

00:29:08: It's an interpreter between humans which I think, you know like going back into the philosophy thing is very interesting because it's this type of interpreter.

00:29:17: You want?

00:29:18: Is it going to change the way we interact with other

00:29:21: people?".

00:29:22: I find that fascinating and

00:29:23: a little bit scary... It kind does already right!

00:29:25: The way especially emails are being written is changing drastically through Chatchapiti or other LLMs.

00:29:33: about the speed of this revolution is really interesting that you made, because I think we keep forgetting how short it is that we had these LLMs and agentic systems at our disposal.

00:29:44: And are expecting also like super great performance.

00:29:48: At the same time It's a year or two years three years.

00:29:52: they have been rolled out.

00:29:53: We're taking baby steps but i think This is something that scares people.

00:30:01: the science space?

00:30:02: Of course, did you get any feedback insights from people?

00:30:06: are people scared of AI scientists AI co-scientists.

00:30:10: Are they afraid that there will become obsolete or job will change drastically.

00:30:16: I didn't necessarily get that, i still get quite a lot of signal that people are dismissive.

00:30:22: so if i go back to this scientist what was telling you about?

00:30:24: he's working for in very large prestigious university and now these tools has changed the way it is doing science.

00:30:31: And he was telling me that he had to interact with people who know they're very senior in his University Who haven't even tried yet dismissive.

00:30:41: I worked for an organization where you also had like two very separate part of organizations, some people were really enthusiastic and i think with enthusiasm sometimes...you know?

00:30:50: You have a bit of positive bias.

00:30:52: but then on the other hand we had people that were so negative they didn't want to touch it because there was extremely dismissive.

00:30:59: And again!

00:31:00: The truth is somewhere in between.

00:31:02: these are tools as said developing at an incredible pace.

00:31:06: So yesterday's limitation will be fixed with the next update, now is it working for everything?

00:31:13: No but this definitely going back to what I was saying in the beginning about challenges that we have with a number of diseases and drugs.

00:31:22: always tell people who are dismissive with AI look It's not like the way we do.

00:31:26: drug discovery and development today is perfect.

00:31:28: I mean, i don't want to wait five hundred years To fix everything.

00:31:32: as you said our health system cannot absorb the cost.

00:31:35: so maybe AI Is Not Perfect but So far.

00:31:39: it's Our best shot at addressing real problem that We have Today just continuing to Do things The Way We've Done for the past thirty Years?

00:31:47: It's Not Gonna Solve it!

00:31:48: If You Have Something Else to Offer I'd love to see it and then be my guest, you know.

00:31:53: It will...I'm open to any new way of doing things and solving non-problems where we can all agree on what is the problem?

00:32:00: And What Is The Magnitude Of A Problem Where We Might Disagree How To Solve It?

00:32:03: But What i'm Saying Is That I'm Seeing Signals With AI that Are Quite Positive Still Just Signals but I Haven't Seen Those Signals with Other Approaches.

00:32:30: maybe one sentence extension to the philosophical question when it comes to R&D in general and of course, AI co-scientist.

00:32:39: You were mentioning that we don't understand anymore what's going on under the hood.

00:32:45: We don't know yet but start using and trusting at the same time.

00:32:53: That is kind of scary.

00:32:55: So I think that we need to unpack the question.

00:32:57: Okay, okay...I have an anecdote.

00:32:59: a couple of weeks ago there was one first warning letter from the FDA to company that submitted regulatory document i think it was a QC document where they said what were done to do the QC for another product and this thing straight out of chat GPT complete hallucination.

00:33:15: The FDA had tell the company look you're responsible for what's your submitting.

00:33:20: You can use any tool you want, but eventually over one signing it and you have a one responsible.

00:33:25: so if you submit something better be something that is real not hallucination.

00:33:30: So I mean eventually we cannot offload completely our responsibility to a robotic system or an AI system.

00:33:38: in the end We don't have AGI yet whatever is proposed.

00:33:47: going back to the human in-the-loop type of question is where we are.

00:33:52: Where there's a proposal or conclusion, it's ours to validate and test.

00:33:59: but you're raising an important point about transparency.

00:34:03: how much we understand about the algorithm.

00:34:05: There were lots companies that tried to provide transparency on what was happening under the hood.

00:34:12: I was at a conference recently taking the kind of black box algorithm and trying to derive the rules from the Black Box Algorithm, To be able say when there is this prediction.

00:34:26: these are things that were taken into consideration by the algorithm.

00:34:31: So it was degrading a little bit quality but made much more transparent.

00:34:38: So that's one approach of going forward.

00:34:40: And again, if you compare it to the early days of chat GPT when you asked a question and then he gave you an answer.

00:34:45: now we still have also another thing... You know?

00:34:48: You can follow how our system is thinking.

00:34:51: It's not perfect but I think everyone agree it's a challenge.

00:34:55: We still want To Have A Feeling That We Are In Control.

00:34:59: so we Still Want To Be able Open The Hood And See How It Works Underneath Now.

00:35:04: If Think About Ten Years Or Twenty Years From Now Maybe you're good with mechanics.

00:35:09: I have no idea how my car function and i'm absolutely fine with that.

00:35:12: so maybe in twenty years we'll be like, You know We trust those thing enough That we don't necessarily need to understand exactly what's happening?

00:35:18: We are still in the early days.

00:35:19: So we are still like...I Like you but I don't trust you yet.

00:35:25: Tell me What your doing ?

00:35:26: I

00:35:27: think there is one other part of it And not all AI Is deep learning rule-based.

00:35:35: AI will not hallucinate in the same way.

00:35:38: So I think we also go back to, or are already at a stage where we have hybrid applications.

00:35:44: so for example QC you might not necessarily use a deep learning model really depends on your use case and some just want absolute certainty.

00:35:55: then you may need different AI models with more room for error but as you say this is an example of self driving car.

00:36:03: You're very forgiven to a human driver who makes many mistakes and might be too old, too tired, too drunk or blind.

00:36:10: To whatever...to drive!

00:36:11: And we are like yeah but that's the humans.

00:36:12: they know what their doing?

00:36:14: Then you have self-driving car which is proven much less mistake than a human drivers would do.

00:36:20: But because it's a robotic system and I mean, because of the legal issues involved.

00:36:24: We don't trust this system.

00:36:26: in the same way And as we learn to accept that AI systems probably will not forever be large language models There are different tools emerging with more world knowledge That have more understanding It is clear situation right?

00:36:46: more reasoning capacity to have more understanding of how things really work instead.

00:36:51: Of just hallucinating them together in an almost perfect way, but I think we will learn To know when to trust these systems.

00:36:59: and went not to trust him.

00:37:00: And i really like it.

00:37:01: We had that in the past episode.

00:37:02: this quote from Ben Evans That having a genetic AI or having AI is having like infinite interns for your work?

00:37:09: Yeah!

00:37:09: And This Is...I Think The Perfect Analogy You Have Like A System That Does Something That's Pretty Good And maybe it's not perfect, but I mean still helps you.

00:37:20: You just need to decide when can trust that and

00:37:23: whatnot?

00:37:23: Yeah!

00:37:24: I agree with what you said.

00:37:25: we are way more forgiving than our colleagues in AI systems.

00:37:30: The number of times i had this discussion where people were telling me they don't like it is not perfect.

00:37:35: the thing was telling them well... Are your colleagues absolutely perfect ?

00:37:41: much higher expectation for an AI system than what we have for humans.

00:37:46: So I fully agree with you.

00:37:47: How would say, where do you see the future trends also when it comes to automation?

00:37:54: AI scientists and discovery?

00:37:56: Where did you see biggest impact on big challenges?

00:38:00: There are a number.

00:38:01: if you look at the whole pharma workflow Interestingly We still don't have lot of new targets.

00:38:07: You know told about those twenty four thousand diseases.

00:38:11: I don't know if i have my number right or not, but they are approximately you know twenty thousand targets and we keep making drugs for like seven hundred.

00:38:21: So what about the other one?

00:38:24: Still waiting new innovative targets.

00:38:28: out of all those targets people will say there's things that are so-called undruggable target.

00:38:33: But this is also where can we help with these systems You know, that don't have like those deep pockets.

00:38:40: That are easy to drug so you can redevelop more molecular glue or come up with new mechanism and new mode of action we never thought about.

00:38:47: I think thats where i'm quite hopeful for AI To find a new target We haven't explored And new ways of interacting With us targets Of course better understanding the mechanism of disease as well.

00:39:00: There is also one thing Thats kind my pet peeve When I look at how we develop drugs.

00:39:04: It's still very much One disease, one drug which I think is extremely naive.

00:39:09: because if you look at a map of metabolism and it's so complex.

00:39:14: You know?

00:39:14: You have things going in every direction.

00:39:16: So If you block something It's like...it's like in the city Like..if we had road then you blocked this road Doesn't necessarily block that whole city.

00:39:23: You will have escaped roads.

00:39:24: And i think its same for diseases.

00:39:26: So I told to a lot of scientists who say Look That might be a direction where AI can really help is thinking about those treatment that will be way more complex where you have not only multiple drugs, but we also think about how do you time it so?

00:39:41: That You Really understand the disease progression and How to block The different avenues of a disease.

00:39:46: So that's one thing.

00:39:48: But the other big Thing Is the acceleration of clinical trials.

00:39:51: So, the FDA now is experimenting with not the sequential way of doing a clinical trial which you do years of discovery then you start the clinical trial in different phase and that takes also years but they are doing continuous monitoring In many ways.

00:40:06: what happened during Covid where we were able to really squeeze things to maximum because we had no choice bringing back this way of thinking taking calculated risk, you know if we can do that with a combination of changing how we evolve the regulation and how we bring AI.

00:40:26: I think these are things also very interesting avenues.

00:40:40: your approach which is fascinating promising from the scientific point-of-view most definitely gets into Yeah.

00:40:53: but when you think about it, I mean the blockbuster is necessary.

00:40:57: When do have something that takes twelve years and costs six billion?

00:41:00: Of course!

00:41:01: If there's some thing like I'm exaggerating just for a sake of argument instead of ten years if it take one year to cost hundred million then you can think completely different.

00:41:12: business model which also very interesting.

00:41:15: I'm not so much into the commercial part of pharma and business model but as science is evolving, that will have an impact on our models.

00:41:24: We talked at the beginning about prevention versus treatment.

00:41:29: You involve a model with more prevention, you have treatments which are more personalized or complex.

00:41:37: it's a complete revolution of the whole healthcare system.

00:41:40: So you know, like that so big and I would not...I certainly have an opinion.

00:41:45: i don't necessarily am NOT necessary an expert!

00:41:47: I don't know if there are a lot of experts who can talk about R&D clinical commercial health care?

00:41:53: That is a big space yeah.. The headline for this podcast is the Bio Revolution Podcast And This Is Where We're At.

00:42:00: Yeah

00:42:00: I'll ask again where?

00:42:01: what do say other challenges ?

00:42:03: Is It Like Data, Model, People, Infrastructure All of it.

00:42:09: I think all the above, i spent many years in the data space making argument that we can't do any other things than what you are doing now if don't have good data.

00:42:18: so still thinking our data is terrible.

00:42:21: why?

00:42:21: because we're not... the pharma industry is a heavy-data industry but life scientists aren't data people.

00:42:28: The way Life Scientist was thinking few years ago and he's still thinking to large extent.

00:42:32: Is there problem?

00:42:34: I'm designing an experiment producing data, analyzing the data and then you know like deriving some conclusion.

00:42:41: And I move on.

00:42:42: it's very linear.

00:42:43: what happens with a data?

00:42:45: You keep them somewhere because either your have to from a regulatory standpoint or Because you know that's what you do.

00:42:52: i mean you keep them in one of your folder but It is very linear.

00:42:55: now we've been forced to rethink things In uh much more circular loop type way.

00:43:01: But then, you know like when we say organizations sit on goldmine of data.

00:43:06: I mean for me this is absolute BS because if you look at the old data...I've looked into archives or large companies.

00:43:14: what does it means to look in legacy data?

00:43:17: Old data going through someone's laptop You open a file and some Excel files any information.

00:43:26: You sometimes don't even have column header, what do you with this Excel file?

00:43:30: Nothing because it's just a bunch of numbers.

00:43:35: so you can't really derive any knowledge from that!

00:43:38: So I think that is where we need to evolve.

00:43:40: and one very strong case i'm always making with people...I am talking too also with clients..is great about the old thing.

00:43:49: but fix forward flow tools that are helping us generate more data.

00:43:56: And I think it's predicted in the next five years, we'll generate more than in the history of humanity!

00:44:02: So you know... It really is an exponential thing and people always say humans aren't good at understanding exponentials.

00:44:09: fix your new data.

00:44:11: Of course, you know there is a value in the old data.

00:44:13: but I think we spend so much money trying to rescue all data that we forgot that We need to fix the new data first because that will be Your New Data Will Be Your Old Data You Know In A Month.

00:44:25: So Fix That!

00:44:26: So i Think That Answer Part of Your Questions.

00:44:28: The Data Foundation Still Thinks That We Are Far From Ideal And Because My Professional Background Most Of my Network Is In Data The interesting thing is that people still have a hard time making an argument, they need to get funding... ...to create good data.

00:44:45: Why?

00:44:46: Because data are not sexy!

00:44:47: You know like people want to spend money and have an AI application.

00:44:51: you know everyone gives money for AI applications.

00:44:53: so now it's interesting.

00:44:55: there is this concept of fair data no findable accessible interoperable blah blah.

00:45:00: So I was talking to a very senior IT person recently, and he was telling me we don't call them fair data anymore.

00:45:07: We called them AI-ready data.

00:45:08: It's the same thing but we get more money for that.

00:45:11: so Data foundation is one thing.

00:45:14: then The other things people People are the main thing.

00:45:16: why?

00:45:16: because?

00:45:18: technical problem eventually will solve it.

00:45:20: But when how people work together with mentality changing your way of doing things That much more difficult thing And its interesting.

00:45:30: As I'm getting older, the thing that fascinates me most is how you get people to work together.

00:45:37: I worked in an organization years ago where we had an AI team full of engineers and then we had medicinal chemists who hated each other.

00:45:46: The AI will send us a list of molecules coming out from their algorithm.

00:45:51: They'll look at it and say My intuition tells me it's wrong, and they don't know what we are talking about anyway.

00:45:59: Then the engineer will say I know those old school medicinal chemists.

00:46:02: you know... They don't live with their time!

00:46:04: And this was really about getting people in a room making sure that they realize there is not an enemy working on solving the same problem.

00:46:12: And that's going on in every organization.

00:46:15: So solving this problem is one of the number-one thing, you know if we want to get people working together against a common goal?

00:46:21: I

00:46:21: think it's also question of the subject matter.

00:46:23: experts who would have their juridic knowledge only they can recreate and then usually are not very open for new things because i think its kind your business model as person To be the only person who knows stuff.

00:46:38: But to your point also, there is of course a fear or you know am I going to be replaced?

00:46:43: And that's something we should not discount.

00:46:44: Of course it is this fear.

00:46:46: but then how do we bring people along in their journey and... ...I think the biggest quality that will need moving forward are these people can operate in different universe ,in different space a strong subject matter expertise but have good feeling and understanding of the other domain, likewise in the other direction.

00:47:05: And this is what type.

00:47:06: people that are open enough to try new things ,to understand, appreciate things which aren't perfect or to be able to co-op with imperfection... That's all we need now.

00:47:18: because when you're on these transition phases between old ways of doing things as well as the new way people who can deal with messiness.

00:47:27: And I think one thing that i'm really hearing and also feeling from many things we discussed about AI, We need to find a way to work because there is no way back it will not go away again... work systems, our processes to include it because otherwise.

00:47:49: I mean one of the things that really triggers me is when i start to see you know organization or company or whatever saying we are banning AI or blocking AI and I think this is so stupid.

00:48:02: Really?

00:48:02: You know whether you like it or not AI's here.

00:48:07: its'nt gonna go away understanding, experimenting not being an unrealistic enthusiast but actually being a realist and saying look it's here.

00:48:18: It is imperfect!

00:48:19: But we need to live with our time.

00:48:20: I mean its like seeing how you are using internet or that tool.

00:48:26: when i see for instance in school they want ban the use of LLMs.

00:48:31: The problem is not banning LLM, the problem is a stupid use of the tool.

00:48:36: Or and I keep telling people look you know kids will use it anyway.

00:48:39: so when you interrogate them don't just read what they bring but tell I don't really care if they wrote it or not.

00:48:47: Did we actually understand what the road can be explained without looking at?

00:48:51: And

00:48:52: there is also, again going back to often teachers are further behind than kids when it comes to use of AI.

00:49:01: so... They're scared too.

00:49:04: have this tool that they don't understand but...

00:49:07: You can blame them!

00:49:08: It's like no one was trained.

00:49:11: We were in a huge societal transformation When no one was ready for that, No One Was Educated For That.

00:49:20: So Some Of The Best Organizations Now Are Putting The Right Programs In Place.

00:49:25: But You Know Going Back To The School System I Remember When Chad GPT Was Launched And There Were A Bunch of Countries You Know.

00:49:31: Banning It.

00:49:32: You Know I Think Italy Bandied Australia As Well.

00:49:35: the One Country Who Said We Will Have Chad GPt in Our Curriculum it Was Singapore Which is interesting, because Singapore is regarded as one of the best places when it comes to education.

00:49:46: And I was like well that makes sense!

00:49:48: We understood this here and we need to treat these as a change which now is part of our life for kids.

00:49:57: so let's make sure they understand the tool... ...and operate with them in an intelligent

00:50:03: way.".

00:50:08: could go on for another like two, three or four hours.

00:50:11: But we're going to re-invite you.

00:50:13: most definitely I think.

00:50:16: let's see about our predictions and absolutely always nice to go back

00:50:21: when we were taking out.

00:50:22: everybody including teachers for that matter have to leave their natural habitats yeah And We Have To Become Friends With Messi.

00:50:30: all of us be it scientists Be It People In Public Service Be It Citizens Right.

00:50:36: That'S a very fascinating takeout.

00:50:39: New business models, new scientific approaches when it comes to drug development – this is a fantastic overwhelming journey I would say what we've been discussing and more ahead!

00:50:54: This was the BioRevolution podcast.

00:50:56: Thibaut thank you.

00:50:57: Well thanks for having me.

00:50:58: Thank You Izzy.

00:50:59: Many Thanks To The Both Of You.

00:51:00: Please

00:51:01: make sure like us at Spotify or Apple or any other podcast platform of your liking.

00:51:07: you may want to check out our liner notes and want to visit our homepage at

00:51:15: science-tales.com.

00:51:18: Thank You, see around next time!

00:51:19: Bye bye.