- il y a 2 jours
AI-Powered Drug Innovation: From Discovery to Development
Catégorie
🤖
TechnologieTranscription
00:01Okay, good morning everyone. It's lovely to see so many faces in our audience.
00:06Welcome to the Blue Stage where we're discussing everything health and tech related.
00:11Now in this first panel with my amazing panelists who I'll introduce very, very shortly,
00:15we're going to be discussing how AI can really transform the process of drug discovery and development.
00:21For too long it's been labored, it's been expensive and with a high risk of failure.
00:26What too is the role of trust and regulation? Well, let's ask our panelists.
00:30So going down the line, we have Jackie Abitol, we have Matilda Strom is on the end over there.
00:37We also have Nina Mian in the middle here and then Dr. Nicolas Lopez Carranza.
00:43So I'll allow all of you to go down the line to introduce yourselves to the audience so they can
00:47know what you do.
00:48Please go ahead Jackie.
00:49Hello, good morning everyone. My name is Jackie. I'm part of a fund called Cathay Innovation.
00:54It's a VC fund, part of a larger group called Cathay Capital, which started 20 years ago.
01:01So Cathay Innovation is a VC investing across three geographies, Europe, US and Asia.
01:08So I'm a partner, I'm leading the European activities.
01:10And we focus in four main verticals, digital health being one of them.
01:16And we just recently announced the closing our third fund. It's a 1 billion funds to invest mainly in AI
01:23companies.
01:24And as I said before, digital health is one of the key verticals. Happy to be here today.
01:31Please go ahead Nina.
01:32Hi, morning everybody. Thank you for the invitation to be here today.
01:37My name is Nina. I work for AstraZeneca, which is a global biopharmaceutical company which operates across oncology, biopharmaceuticals and
01:48rare diseases.
01:49My work specifically focuses on the biopharmaceuticals area, where we're working to transform the lives of billions of people that
01:58suffer from chronic diseases.
02:00So to that end, I lead a data science team.
02:03We operate right from discovery through to development, where we are really looking at supporting decision.
02:15So decision support providing everything from rapid insights from data through to new systems and solutions,
02:25which overall enhance our capabilities and our ability to do research and development.
02:32is good.
02:33Yeah. Good morning, everyone. It's a pleasure to be here. Pleasure to be at VivaTech.
02:37My name is Nicolas Lopez Carranza. I am the head of BioAI team at InstaDeep.
02:44InstaDeep is a startup. It was a startup that focused on decision making for the enterprise.
02:51It has a few verticals, still is serving other customers.
02:56But in 2023, due to a very successful journey in the domain of BioAI, we got acquired by a large
03:06pharma group.
03:06Not as large as AstraZeneca, but it's BioNTech.
03:10And now I focus my work on developing AI research and applying AI models for BioNTech pipelines.
03:20Yeah. Pleasure to be here.
03:22And then Matilda, please go ahead and introduce yourself.
03:25Hi, everyone. My name is Matilda. I'm the COO of a company called Bioptimus.
03:31We are building the world's first multi-scale, multi-modal foundation model for biology.
03:38So we are tokenizing all of the languages of biology in all of its scales and modalities so that we
03:47can simulate a cell, simulate a tissue and eventually simulate a patient.
03:52I'll go into a little bit more detail what this all means later in my answers.
03:56Thank you so much, panelists. And definitely we will discuss that because it is fascinating and you'll learn a little
04:01bit more about.
04:02But first, Jackie, I want to come to you then. As an investor, how exciting a moment is this right
04:08now with the use of AI to help drug discovery?
04:12And perhaps you can paint the picture of where we were when it came to drug discovery before AI.
04:19Okay, sure. So maybe I will start sharing an anecdote about when we did our first investment in healthcare.
04:28Okay, when we started with fund one, so 10 years ago, and we were, you know, fundraising and we're saying,
04:33okay, we're sector agnostic, but we will do everything but health.
04:37So it's funny because 10 years after where I'm sitting here with just, you know, people coming from an industry
04:42that 10 years ago, actually technology was not that impactful.
04:47And our first investment came because we met these two co-founders of a company called Hawkin. I think it's
04:52a well known company, a French based company, where you had the combination of oncology and tech.
05:00Okay. And we realized at that moment, the impact of technology in basically reshaping pharmaceutical companies from drug discovery, but
05:10to many, many other things.
05:12And we said, okay, maybe this is the right moment for us to be an investor, to be active in
05:17that space. I'm talking about it was in 2017. So eight years ago.
05:23Coming back to your question, what we've seen is especially since, I would say since the COVID outbreak, I mean,
05:29we're seeing the emergence of, you know, AI reshaping.
05:32I mean, it was not called AI, it was called ML, it was called many, many things, but not, not
05:36what we're seeing today, especially, I would say that the last couple of years from 2023 to 2025, we are
05:43seeing how AI is not only accelerating timelines, improving accuracy,
05:48and basically, as I said, reshaping the whole, the whole industry in drug discovery. We, we've been monitoring the space,
05:57I think since 2020.
06:00Okay. When we come across a company, also a French company called Alkimia, and I'm not here to brag and
06:05to talk about the portfolio, but I think it's quite relevant to your question.
06:09You can talk a little bit about the portfolio. Go ahead, go ahead.
06:11No, but it answers some of your questions. And so basically, they resolve an equation. So they're using quantum physics
06:18and AI to help drug discovery, to accelerate.
06:22And what's really important is not only what the company does, but it's also how the industry and how players
06:28like AstraZeneca or Sanofi, for instance, are now partnering with this type of companies to accelerate their processes.
06:37So basically, making sure so that the whole from the identification of the pipeline to the test to the in
06:44silico.
06:44I mean, I'm sure that Nina will tell us a little bit more about how a drug comes into the
06:49market.
06:50So using using AI, and I think it's not only about the tech, but it's also the market adoption is
06:55here.
06:56And we are feeling that this is the best moment. I mean, to your point, the best moment as investors,
07:01but also as the industry.
07:04So to have this combination, which maybe 10 years ago was like, you know, this sounds like crazy, but now
07:11it is happening.
07:12Well, you teed up Nina perfectly then. So I mean, he ended on 10 years ago. It was crazy.
07:18Is that how the pharmaceutical industry has seen it as well?
07:22Yeah, absolutely. I mean, it's such a important time in terms of how we're thinking about the use of AI
07:30to understand and treat and find molecules for disease.
07:35The burden of disease in the health care system that's impacting people in society is is ever increasing and how
07:45we're handling it today is just not sustainable.
07:48So at AstraZeneca, we are using AI from R&D through to operations.
07:54If we think specifically in the R&D space, AI is already enabling us to do things much faster with
08:04much greater scale and uncovering insights that frankly just weren't possible before.
08:10So if we think about perhaps the diagnosis and trial space specifically, we are also using multimodal AI where we're
08:23integrating pathology and imaging and real world data to generate clinical decision support tools in the oncology arena, specifically a
08:36recent example in breast cancer.
08:38We have a tool called Milton, which again, integrating across multiple different data sources, so genetics, proteomics, and real world
08:53data again, to much earlier on understand how things are changing in the body to detect disease, those signals of
09:04disease much, much earlier on before an actual diagnosis is made.
09:07And so Milton can do that across more than 1000 diseases.
09:12Finally, in the trial space, we're using AI to look and understand who is much more likely to respond to
09:20disease so that we are not exposing people to molecules that they shouldn't be exposed to.
09:27But also using AI techniques to extrapolate and interpolate.
09:32So perhaps where a biomarker hasn't already been measured in the system, we have an algorithmic patient selection suite of
09:41tools which allows us to infer what that biomarker might be and allow clinicians to make much more informed decisions
09:48as to who they would invite for pre-screening to then join a study.
09:52So those are just a few examples of how we're integrating AI across multiple parts of the R&D spectrum.
10:00Now you mentioned there burden, but also risk. So let's pick up on both those things.
10:04When you mentioned risk, so that's obviously, you know, the benefits you could say of AI right now.
10:09But when we come to burden, you know, obviously hate to mention it.
10:13But of course, we've just come off the back of a pandemic and we've seen, you know, the scale of
10:18which, you know, health systems have been completely devastated.
10:20So do you think AI is coming in or you could say, you know, the start of the revolution of
10:26the AI, you know, the multi use of it?
10:28Is that now starting at the right time, given that, you know, the entire world has been through this massive
10:32crisis as well?
10:35Well, I think there's a lot of combining factors, right?
10:38So the pandemic has certainly accelerated the use of technology.
10:42So if you think about things like telemedicine pre pandemic, that was something that I think a lot of health
10:48care systems were aspiring to be able to do, but they were still trying to figure out what's the reimbursement
10:53model specifically in the US.
10:56But through the pandemic, many of these things, you know, in clinic visits just weren't possible.
11:02So in terms of the acceleration and adoption of digital technologies and digital often is the foundation for AI because
11:10there's a lot of data that goes along with it.
11:12And with that, we've seen the pandemic has been a leveler in that in that sense.
11:17And so Nicholas and also Matilda, then take us through how I'll start with Nicholas first, how from, you know,
11:24AI has really revolutionized your processes and if you can give some examples.
11:28Yeah, absolutely.
11:29And it's interesting, Nina, you said you have been integrating AI into the biology pipeline, the pharma pipeline.
11:36We have been born as an AI company, so AI first company, and we were integrating biology.
11:43That was our journey is, okay, we have these amazing models, reinforcement learning, foundation models like NLP, how can we
11:51apply them in biology?
11:52And we were one of the first actually doing that.
11:55A few examples, we developed early nucleotide transformer.
12:00Today is the foundational model for genomics.
12:02It's one of the most cited models. We published it on Nature Methods.
12:05We have more than 300 citations on this model.
12:08And it's a benchmark that if you are serious about genomics and AI today, you need to benchmark against nucleotide
12:14transformer.
12:15This is like what EVO2 and other models do.
12:19And we have also been working on the proteomics field where, you know, classically to sequence a peptide, you would
12:28use a database.
12:29Now what you call a target decoy search.
12:32So you know the tandem mass spectrometry, you have your spectra, you need to find out which peptide is out
12:36there.
12:38So you see if the spectra matches what you are looking for in your database.
12:43However, now with AI, with sequence to sequence models, the same models that would translate from German to English,
12:50you can make them translate from spectra to actually a peptide.
12:56And this is what we did with InstaNovo.
12:58Again, we published it in the Nature Journal.
13:01And it's quite a state of the art today on peptide calling.
13:04And yeah, we open source it.
13:07So the community is using it and is giving us great feedback on this.
13:11It's quite exciting to see how we combine all these things.
13:13A bit like what Matilda and actually everyone here is suggesting.
13:18So these are the things that excite me the most is how do we take all these separate models, bring
13:24them together
13:25and bring solutions to the customer and predict phenotype in particular.
13:32Matilda?
13:34I think your question was essentially what breakthroughs are we seeing or do we expect to see?
13:39So, of course, I'm going to say that foundation models are already showing that they're going to be the ultimate
13:45breakthrough for AI in biology.
13:48But let me start by backing up and understanding what AI has done already in the drug development and drug
13:54discovery space.
13:55So, to date, 21 AI-designed drugs have made it through phase one trials.
14:01And a couple of AI-designed drugs have made it through phase two.
14:05In fact, one very recently from in silico medicine.
14:08This is about two or three percent of all of the drugs that get through every year phase one.
14:15But what's interesting is that the AI-designed drugs tend to have been more successful in their phase one.
14:23So about 80, 90 percent success rate versus anywhere from 50 to 70 percent success rate for non-AI drugs.
14:29So already there is starting to be proof that AI can start to drive better results so that you have
14:35fewer trials and more efficiency and more accuracy in what you're doing.
14:40But, so far, the AI that's been developed has largely been developed for single tasks.
14:48Whereas foundation models that are by design built on unlabeled data to solve a variety of tasks.
14:56Think OpenAI and ChatGPT.
14:58They weren't designed for a specific task, but rather to answer any question.
15:03Foundation models are now coming in with the advancements that we've started to see in some of the technologies to
15:09capture data.
15:10That there is now enough interesting and multimodal data out there that's large enough in its capacity that foundation models
15:17can now come in and drive very interesting insights.
15:21But, so far, the foundation models have still followed the traditional track of biology, which is to stay in its
15:27silos.
15:27So, depth in protein like AlphaFold, for example, which is fantastic.
15:32And I think there's even, you know, CZI and Recursion have recently put out amazingly ambitious papers around creating a
15:39virtual cell.
15:41But what we need to see is the ability to translate across scales.
15:47So, basically understand how that protein interacts in the cell and then interacts in the tissue and then within the
15:54whole human body.
15:55So that we can understand patient response better, toxicity.
15:59Maybe there's multiple drug targets that will solve a problem rather than just one drug target.
16:06And if you have a foundation model that's looking at all the different modalities, you're going to be able to
16:11find insights like that that are really not catered to the way things are working today.
16:16And I think we're just at the cusp of seeing how this is going to impact the whole space.
16:21And I think it's going to take a lot because I think the whole industry is really geared towards this
16:26linear generation of outcomes.
16:28Both, I mean, pharma, pardon me, Nina, and regulations.
16:33And so, you know, where foundation models can play at every level simultaneously and instantaneously, where it might have taken
16:41years for each stage.
16:43You know, we're going to have to unpick a lot in order for this to work.
16:46But foundation models really are starting to show that there's something that can be blown out of the water here.
16:52Nina, I can see you nodding a lot. Your thoughts?
16:55Sorry, I missed the question.
16:56I see you nodding a lot in agreement there.
16:59Yeah, yeah, absolutely.
17:00I mean, I think there's a dual track here, isn't there, between the evolution and the transformation.
17:06So AI has got a role to play to accelerate and innovate in how we're currently operating.
17:14But you're hearing all of these possibilities and all of the potential, and we're really just at the beginning of
17:19that.
17:19And essentially, I think we are going to see completely new paradigms of how we operate and how we understand
17:27biology.
17:28And it's an incredibly exciting time.
17:30It is.
17:31And that sort of teased up, Matilda, then.
17:33I mean, you were talking about foundation models.
17:35So talk us through your foundation model for biology.
17:37And if I may, I think you've said that it will be able to take any conceivable health question and
17:43give the person the best answer.
17:45But how do you then stop something like that, not becoming Google Doc?
17:50So when the average user thinks, oh, I've got a small problem, and then they go on the internet and
17:56they sort of end up having or believing they've got an even bigger problem going on.
18:00So how do you stop that?
18:01But talk us through your model.
18:04So our model, as I mentioned, is we're building it up from the start to be AI and biology from
18:09the beginning, right?
18:10I think most of the companies have come from InstaDeep, I suppose, a kind of direction where they start with
18:16AI and then try to apply biology.
18:17We're trying to do both at the same time and start top down to understand the interactions within your body.
18:24And what that means is we're taking multiple modalities, so that means things like H&E slides, spatial transcriptomics, IHC,
18:35and building up to patient level and clinical health records
18:39so that we really have an understanding across the whole patient's biology.
18:46So that's what we're building, ultimately.
18:49And what it means, you know, you're asking what happens if, can this model really be able to answer any
18:55question conceivable?
18:58I think we're very far from that ultimate vision and from, you know, a user using it like search, but
19:03potentially our grandchildren will be able to take a biopsy or a blood test, put it into a model,
19:08and be able to understand how they will react to a certain, they specifically will react to a certain drug
19:14or treatment.
19:16I wouldn't very quickly put it into the hands of a layman.
19:20I think we have to start somewhere.
19:22And where we're starting is you still need to, you know, the difference potentially between building an LLM foundation model,
19:30where you can take any text and that you can learn and tokenize texts and understand contextually how to answer
19:37questions.
19:37I think in biology, you still need to train the model on some certain specialisms and outcomes because you need
19:46the data that helps it to learn what happens if then.
19:51And that's really specific to diseases.
19:55It's specific to the cells or proteins.
19:57And so you still need to build it sort of stepwise in a sense of teaching it how to do
20:03things within one realm first and then build upon that and then build upon that.
20:07And then I think we'll get to a stage where it can answer almost any question.
20:12But largely, it will be not in the hands of your individual people for the foreseeable.
20:19But how far away do you think that could be then?
20:22Because obviously, you know, this is a stage where we're talking about, you know, health and technology.
20:27So for the average person who wants to have, you know, understand their health a bit better.
20:34How long would it take for your technology to get into their hands?
20:37Do you think you had an estimate?
20:39That's a really good question.
20:41I think because of the not just the speed of technology and data that's available to build this, but also
20:48regulatory regimes and people's perceptions.
20:51It's we're probably a decade away from that.
20:55But the ability of the technology, we might be five years away.
21:00So, you know, but I like the question because actually we are also investors in Bio Optimus and I never
21:06asked Matilda that question in the board meetings.
21:08So thank you.
21:09Thanks for asking the question for me.
21:11So I guess my next life is in the boardroom, I guess, then.
21:15Nicholas, you wanted to say something too?
21:16No, no.
21:17What I wanted to say is that like a bunch of AI experts revolutionized biology and got the Nobel Prize
21:25last year, right?
21:26Like the CAS competition for 30 plus years, people were struggling with like, you know, like, you know, molecular dynamics
21:34or protein structure prediction,
21:36using like moving things with bits and pieces of the proteins.
21:39And some people take the whole PDB.
21:41Let's not neglect that the data was there, right?
21:44Like 50,000 protein structures available on PDB that they could train on.
21:50But they smashed completely the CAS competition in Alpha 4.1 and later in Alpha 4.2 where they basically
21:57reach the same accuracy, experimental accuracy, you know?
22:01So here you see a shift of paradigm, you know?
22:04People who don't know how to or are not experts in Go become Go experts.
22:08People who are not experts in biology disrupt completely the field.
22:12And this is where AI is so exciting, right?
22:15Because it comes, the punch comes where you're not expecting it from, right?
22:20Yeah.
22:21So, and in terms of regulation, I totally agree with Matilda.
22:26We are quite far.
22:27I feel that we need to be responsible.
22:31I think that the patient needs to, essentially, we need to do what is right for the patient.
22:38And this is essentially make sure that we are helping the patient.
22:42We are helping cure the patient with these AI models.
22:44And very importantly, we are not damaging them, right?
22:48So we need to work with the regulators on that.
22:51Data privacy is another aspect that is quite important.
22:56I think that pharma industry is quite on top of this.
23:00And moving from a startup kind of environment to now like a pharma group, you know, we could feel that.
23:07And like, you know, we take very seriously data, patient regulations, GXP compliance, quality management systems and all that.
23:17Yeah.
23:17So these are the things that, you know, this is why we are a bit far to actually deploy this
23:24at scale.
23:25But would you say this sort of next sort of building blocks or the things you're very mindful of then
23:30is trust and ethics?
23:33Yeah.
23:33Yeah.
23:33I mean, I think companies are doing it for the ethics first before, like, you know, the regulation.
23:39We really want to make people's life better.
23:42And we are we have an ethical responsibility.
23:45So this is this is grounded on ethics and compliance with regulations, I would say.
23:53And Nina, as a pharmaceutical representative in our panel, then I mean, obviously, ethics and trust is really important for
23:59your business as well.
24:00Yeah, absolutely. So if we think about regulation and how it pertains to the conversation today, there's there's two types
24:08that we predominantly need to think about.
24:11So that's medicines agencies like the EMA here or the FDA in the US and then also AI specific regulation.
24:19But as but setting that aside, even even before the regulatory bodies kind of caught up to the potential of
24:29what I can do,
24:31AstraZeneca was one of the first pharmaceutical companies to publish policies about how we will and how we are using
24:40data and AI.
24:41So that means that when we are putting in place AI solutions, we have committed to make sure that they
24:52are transparent, that they are accountable, that they're fair, they're human centered and socially beneficial and that they're secure and
25:03they're private.
25:04So that means every time AI is deployed, you know, that it's being deployed, you know, that you're interacting with
25:11an AI decision, we take accountability for the output of that.
25:17So, you know, we have quality and checks in the in the system, we know that it's being deployed in
25:24areas that are actually beneficial rather than detrimental.
25:27And that that takes a lot of consideration and judgment. And of course, we have the utmost care in terms
25:35of the data privacy and security. So within AstraZeneca, we have AI governance at an enterprise level.
25:45We also have embedded within the different areas, different functions, so that it can be very specific to the type
25:54of data and AI challenges that are happening there.
25:57And it extends beyond the company. Right. So we we apply this when we're working with third party suppliers and
26:04when we're putting partnerships in place as well.
26:06So when you when you look at it, actually, we we put ourselves in a position where we hold ourselves
26:13to account above what is necessarily nationally and internationally required.
26:20But that is also really important for pharmaceuticals to do that, isn't it?
26:24Yeah, absolutely. I mean, it's it's we need to instill trust, whether that is in the AI space or the
26:32data data space or more broadly. Absolutely.
26:37I wanted to mention something about regarding pharmaceuticals. I think what we've seen as investors is also the shift of
26:46partnering with this type of companies.
26:48And I would like to, you know, we talked a lot about maybe 10, 15 years ago about open innovation
26:53and that time in different industries, including pharmaceuticals,
26:57pharma industry, a bit incumbent players were, you know, partnering with startups just for, you know, making sure that they
27:04would not miss something.
27:05So it's OK, you know, and maybe we'll do a pilot and it will be nice. It will be interesting.
27:09We can learn from it.
27:10And then we move into the digital kind of transformation. But now, AI is really transforming these industries.
27:18And when it comes with, I mean, groups such as AstraZeneca or Sanofi or Novartis, I mean, all the big
27:23names, they are relying more and more into this type of technologies,
27:28because maybe going back to the first question, but just listening to Nina and talking about regulation, talking about many
27:34things.
27:34You can see that everything has been addressed to work with technology companies. And this is for me, one of
27:44the main reasons why the booming is not only happening from a tech point of view,
27:49but also tech, you can find the funding. But if you don't have access to the business, basically, you're stuck
27:54with an amazing technology that nobody will use it.
27:57And what we're seeing is really now this type of, you know, exits and collaboration and breakthrough, which are extremely
28:05important, especially in this industry.
28:07You mentioned collaboration then. So what areas of, you know, AI innovation within the field would you say are still
28:15under invested?
28:19I think. And perhaps where also are you looking to go to next? Yes. No, I mean, the way we
28:26work and the way we look at it, OK, we I mean, we have strategic investors in the different areas
28:32where we invest.
28:33And by strategic, I'm talking about large former groups, OK, without saying any names. And they share with us their
28:39concerns. OK, what is the roadmap? What are their needs?
28:42So the way we look at it is from the customer, I would say, point of view, what is the
28:47pain point? And we are seeing what are the trends that are emerging. OK.
28:51And what we're seeing going back and back, maybe I would say I will focus on a couple of areas
28:56here is clinical trials.
28:58I think AI can play really a very important role in clinical trials because first of all is to make
29:05sure that you address the right population.
29:06So if you just do a clinical test, you have a certain segments of population, which is not representative of,
29:13you know, from an ethnicity point of view, from many other reasons.
29:18And sometimes, OK, the clinical trial fell because basically the patient that you have onboarded in your test are not
29:25the right one. OK.
29:26So AI can help, you know, to do segmentation to identify what are the right patients, which at the end
29:35will help, you know, drugs to go into the market with a higher rate of success.
29:40So this is one. The second one that we are seeing as well is real world evidence, which is very
29:47nice. OK.
29:48But today is a lot of data. I mean, Nicolas was mentioning about a lot of data available, but still
29:55not, I mean, fragmented, OK, or biased or sealed out.
29:58So how do you take this data and use them and work with them even like from foundation models?
30:04And this is something that today it's still not yet completely financed or backed by investors, but it's happening.
30:14It's happening and we're moving from, you know, real world evidence was a lot of, I would say, human work.
30:20And what it means human work, it means that it cannot scale. OK.
30:24But what AI can bring into that, it's scalability, it's automatization.
30:28So and therefore, it's a model that can really work.
30:32So I would say that on our side, probably these are the two sub, I mean, segments or segments of
30:39the industry that we are looking with a lot of interest.
30:41And we are seeing that there's a huge potential for growth.
30:44Maybe I have a comment there or maybe a question. I don't know if it is under invested, but these
30:50AI models are data greedy, right?
30:52We need tons of data. And sometimes when we talk to the labs is, oh, I can offer you two
30:5796 well plates. That's what you have, right?
31:00And this is the challenge, how we get more like deep screen in data, how we get high throughput data,
31:07how we get the lab in the loop.
31:09So I feel that there the industry needs to evolve and that will require a lot of investment.
31:15So my question to you, Jack, is what is the investment these days on data generation for AI specifically?
31:25So, again, I don't want to talk about all the companies that we backed in the space, but we are
31:33seeing.
31:33And actually, for me, the most important part is not under finance. You're right.
31:40Okay. It's basically whenever you go to a customer is how you can use those data because you have tons
31:45of data.
31:46Okay. And we know that the customers are talking about the farmers are in the need of that.
31:50But today, it's mainly project based. Okay. Not a platform based. Okay.
31:55And this is why investors are a bit not afraid, but I would say prudent and cautious because you don't
32:02want very long cycles.
32:04You want, I mean, scalability, but you have companies doing that, but you have companies doing that, doing projects which
32:11can last 12 months, 18 months.
32:14And this is not a model that you can accelerate and you can scale. So that's why this is a
32:19major concern because we know and we feel the pain. Okay.
32:23We know the companies out there, but it's really how you have, I mean, AI can just accelerate the access.
32:30Okay. Of data.
32:31The, I would say the cleaning of the data. So to make sure that we are providing, I mean, all
32:37these companies are providing the right data to the farmers.
32:41Does that answer your question?
32:42Oh, go ahead. Yeah, Matilda.
32:44I was just going to add a point on the data side, because as we said, I think foundation models
32:51need as much data as possible.
32:53They need large and vast amounts. But actually in biology, what it turns out is that quality trumps volume right
33:01now.
33:01If we can get the right quality of data and quality is multimodal data. So, you know, if a patient
33:08is undergoing a treatment, how many different angles can we understand that patient's reaction to that treatment?
33:15And over what period of time can we understand that? Normally you get one single indication, maybe once in that
33:23patient's journey. And if you can get lots of them, you can really understand it.
33:27And that richness of data will then help train the model in the right way to get you the right
33:32answers.
33:32And what's happening with from a data investment perspective that I'm seeing is that it's typically becoming down to the
33:39foundation model players to generate their own data or to find novel ways to create consortiums.
33:45Like Oaken has done with Mosaic, creating a federation of seven different research hospitals to create the largest spatial transcriptomic
33:53atlas across seven modalities, seven cancer types.
33:57That's 7,000 patients. So we're not talking billions, you know, hundreds of billions of words like LLMs are being
34:05trained on.
34:06But that 7,000 patients is still extremely rich and rich enough to give you lots of really interesting insights.
34:12Because maybe I can explain the volume of information that's captured in biology by explaining this from a tokenization perspective.
34:20So if you have what's called a spatial image of a biopsy, so you take a sample of a tissue
34:28that might have a cancer tumor in it,
34:29you take a picture of it with this really great technology that helps you see not only the image on
34:34top,
34:35but down to the cellular and down to the molecular level on this one image.
34:40If you take a six and a half by six and a half millimeters squared section of that image,
34:45that contains approximately 10 million tokens.
34:51So think about that across the human body and how much then a model would need to consume
34:57and how hungry that model would need to be on data, but also how rich some information can be
35:02and what it can start to learn about the human body and about biology.
35:07And one more thing on ethics, sorry, because I missed that topic.
35:11But this is important because I think the diversity aspect is what everyone's very afraid of.
35:16I'm part of this chat group of lots of tech leaders and there was a very emotive discussion about
35:21AI is going to be the worst thing, especially for minority groups who will be underrepresented.
35:28women, ethnic minorities are going to be underrepresented and therefore actually suffer more from AI designed drugs.
35:37And I think therefore companies like ours need to be ethical by design in the sense of the data that
35:43we capture
35:43has to start from a basis of being diverse.
35:46And interestingly, diversity is almost more important.
35:49and we're finding right now to be also on like a technology and a scanner level and a hospital level,
35:55because every annotation even that the clinicians will make in hospitals might be slightly different.
36:02And so you need to make sure that if somebody is treated with the same type of cancer,
36:06even if they have the same ethnic ethical, ethnical background in another hospital,
36:11they may actually sort of the data input may look slightly different and you may get a slightly different outcome.
36:16So there's so much complexity in the data capture here that makes this really important.
36:22But one other thing I would say to what I was trying to argue in this group is there's also
36:27a question of medical breakthroughs, right?
36:30So the same people who may be arguing that, you know, we should not even start the AI journey
36:36because it will automatically backfire on a number of people are also the ones like everyone when they become unhealthy,
36:43who will want to have the latest drug that might help them, right?
36:48And so where do you, where is it, you know, ethically, where do you draw the line to like,
36:53we're going to stop medical breakthroughs and potentially like, you know, some companies do and kind of gate keep their
37:00own data
37:01and say, we're not, we're not going to share this to the world because this is just for us to
37:06create breakthroughs on, you know,
37:08where does the ethical line stop, right?
37:10I think as with everything in the world, you can't be binary on this topic.
37:14You need to then open it up to make sure that we create the breakthroughs that are also going to
37:19be relevant for people.
37:20Maybe I can bounce back on those very relevant topics.
37:25The one in diversity, for example, an example is only 2% on GWAS studies, I was reading an article,
37:36only 2% of African population has been included, right?
37:39In the genetic, you know, information on those GWAS studies, genome-wide association studies.
37:46And the African population is like roughly what, 14% in the world.
37:50And more importantly, from a statistical point of view, because the African genetic is more diverse,
37:57they should be oversampled, not downsampled, right?
37:59So definitely diversity is very important.
38:03Regarding foundation models, I would separate it in two because of course, you know, there is this self-supervised, okay?
38:10You show the internet and you tokenize it so you do mask language modeling, let's say.
38:16And so you don't need labels, right?
38:17You just do masking and you tell the language model, guess the word, guess the next token,
38:22and that's how you train these models.
38:26However, there is also the fine-tuning where you need label data.
38:30And this is where, like, you know, we do need to find solutions that give us those labels,
38:35that give us those phenotype, right?
38:37Because the foundation is useful only if you fine-tune it later for a task, right?
38:42Yeah, so this is what I wanted to bounce there.
38:45And then Nina, just to quickly bring you in on the sort of equity and the diversity aspects
38:49when it comes to data as well, for a pharmaceutical.
38:52Yeah, I think we're really beginning to hit some of the heart of the issues.
38:56You know, you have to have the right data, otherwise the models are never going to be representative.
39:00So I think we've covered that well.
39:02Perhaps what I can add is a couple of additional points.
39:07So absolutely, we need to create those public-private partnerships that mean we have well-represented populations
39:14and work from that.
39:17When you've actually generated the models, though, it's not a one-and-done.
39:21You know, we also need to make sure that those models continue to represent the populations that they're designed to
39:29represent,
39:29so that the models don't drift over time, and that we're looking how, when they're deployed into systems,
39:34they are continually validated and continually updated.
39:38And I think another critical point is to think about where they're used,
39:43really making sure that when you generate a model, if you understand there are limitations,
39:50and there will be because you create a model often for a specific purpose.
39:53I know that, you know, there can be general cases, but you're creating them for a specific purpose.
39:59Whoever's using that model and how it's deployed, you really need to understand that,
40:04so that those models don't then get used in ways that they're just not intended to be used,
40:11because you will end up with bad consequences, you know, so it's about educating.
40:15So it's also about attracting the right people to the field as well.
40:18Yeah, it's absolutely about making sure that the models continue to do what they're supposed to do,
40:23testing them along the way, and making sure that as models become much more pervasive in healthcare systems,
40:29that people understand what they're doing, how they're using them, where it's appropriate and where it's not.
40:36Okay, so I can see that we're at sort of just over four minutes left.
40:40So I'll let you all answer a final question, which would be, you know, looking ahead to the next five
40:45to ten years.
40:45I mean, it seems so far away. Even five years seems like a long, long time.
40:49But if you could perhaps chart, look to the future over the next five to ten years,
40:54what's the one breakthrough that you would like to see happen, either done by yourself,
41:00or generally, what's the one AI breakthrough you want to see happen in the world?
41:05Go down the line, please. Jackie, first.
41:07I mean, first of all, expecting Mathilda.
41:11Mathilda, no pressure.
41:13No, more seriously, I think when we look at mobility, okay, what we are seeing now happening is autonomous driving,
41:24okay, self-driving cars.
41:25So I believe, and we talked about diversity, and we talked about minority, and we talked about ethnicity.
41:30And actually, what I would like to see, okay, it's basically an autonomous drug discovery design and development platform, okay,
41:40where basically everything has been conceived by AI, from the design to the clinical trials, to the, I mean, to
41:48bring the molecules into the market for exactly what you've been saying, okay,
41:53to make sure that we address the whole population, and not only, you know, whatever portion of the population.
42:02So that will enable any player from pharmaceutical, but also biotech firms.
42:07So basically, to even find a solution for a niche or for a rare disease, okay?
42:13So that will be, for me, like a major breakthrough.
42:16Are we there? Not so sure, okay?
42:18It will happen within the next five to ten years, to your question, not so sure.
42:22But I would say that that is probably the end goal.
42:24So to make sure that something like that could happen, I mean, there is the hurdle of cost, basically, okay?
42:32And that can reduce massively the cost, talking about 60% to 70% cost reduction in R&D.
42:39So you can have, of course, the barriers from the people.
42:42I mean, what Matilde was saying, okay, but people are just saying, okay, but technology will replace, I mean, personal
42:49workers.
42:52The skills, I mean, automatizing, even a lab.
42:55And I don't want to take too much of the time, so I will pause here.
42:58But that will be, for me, probably something that I would love to see.
43:01Well, I hope you can do it, Nina.
43:04It's such a difficult question.
43:06I can't predict what's going to happen in 12 months, let alone 10 years.
43:09Okay, six months, then maybe.
43:11What's the one breakthrough that's coming through?
43:12No, I think what I would say is that everything is moving towards agentic systems, right?
43:17So at some point, we are all going to be managers of our own agentic workforce.
43:24And I think what we have to do as well is really prepare ourselves for the fact that we are
43:30in for an unpredictable future.
43:32So learning to move quickly and adapt, learning to be open and think about new business models, because things will
43:41be radically different.
43:43And really, if we can put ourselves in that position where we are open to embrace things, the limit is
43:51our own imagination.
43:52Really nicely said.
43:53Nicholas.
43:55Yeah, I mean, agentic sounds very exciting, so I will try to bring another axis here.
44:00But what I can say is that, regarding agentic, is that at least in the pharma industry, AI will not
44:07replace humans anytime soon.
44:09But maybe a human using AI will displace humans.
44:14So competition will be tough and we need to be ready, right?
44:17That's what people say.
44:19And to bring another axis, as I said, I am very excited of how AI can bring personalized medicine.
44:27Like, you know, understanding the patient, understanding the patient end to end, from the genomics to the histology to everywhere,
44:33taking all of the data, all of the electronic health record, what is going to happen there,
44:38and how personalized medicine led of AI will happen.
44:43That's very exciting for me.
44:45Yeah. And then Matilda.
44:46It's hard to go last, because I was going to say pretty much all of those things.
44:49No clinical trials, right treatment for the right patient.
44:54But what that probably means is that we should be able to find the cure for many cancers in most
45:03people.
45:04Amen. Seriously. Let's hope that, yeah.
45:08That's it.
45:08That's it. Okay.
45:09Well, listen, thank you all, great panelists.
45:12You know, Jackie, Nina, Nicholas, and Matilda.
45:15Round of applause for them, please.
45:16We're going to have our next session, AI versus aging, so don't go anywhere.
45:20Thank you, hopefully, everybody.
45:21Thank you.
45:21Thank you.
45:21Thank you very much.
45:22Thank you.
Commentaires