- 5 hours ago
AI, quantum simulation, and synthetic biology are accelerating drug discovery in ways that were impossible just a few years ago - compressing timelines, expanding what’s scientifically feasible, and opening new paths for treating unmet medical needs. As these technologies move from research to real-world impact, regulators, developers, and the public face new questions around evaluation, trust, and access. This panel explores how we can modernize oversight, build confidence in AI-enabled science, and ensure that faster innovation translates into broader patient benefit.
Category
🤖
TechTranscript
00:22Good afternoon everyone, welcome to a session, speeding past safety, how AI and
00:29quantum tools are reshaping drug development. Very excited to have you guys with us to touch
00:35on this this afternoon. Drug development takes time, sometimes 10 to 15 years, billions and
00:41billions of dollars just to get a single treatment to market. AI promises to change that. We are seeing
00:48some acceleration in discovery, in trial design, in how researchers understand disease. But there's
00:54a lot of tension underneath. The faster you move, the more you need to ensure people's safety.
00:59And the talk about AI and quantum, it makes it harder to tell really what's working today and
01:06what's going to be useful in the future. So that's what we're here to dig into. Having an unbelievable
01:12panel with us today. Very, very exciting. So maybe you can introduce yourselves. Akilah,
01:18we can start with you, please. Absolutely. I'm Akilah Kosaraju. I'm the CEO and co-founder of Fairbio.
01:24At Fairbio, we're using AI and most recently generative AI to design novel antibiotics to
01:31outpace the looming crisis in antibiotic resistance. AI happens to be a really complementary tool
01:38for the problem we're trying to solve, which is really developing novel antibiotics that will
01:43not recapitulate resistance. Thank you. Yes, I'm Tautar, Global Head of Data and AI at Sanofi.
01:53And yes, we are biopharma company. And our goal is to bring the treatment as fast as possible to the
02:00market. Today, it's 15 years from discovery to therapy. So we are using AI on the research side of the
02:08molecule clinical trial, patient recruitment on the development of the manufacturing, but also on the
02:14manufacturing supply, yield optimization, resource allocation, but also on the commercial side for
02:21sales rep to improve also the best next action on the market. And of course, we can do that without
02:29the
02:30employee, without human. So we are really using AI at scale for us as a people in the company with
02:38a tool
02:38that we call concierge. It's like the hotel staff to do everything for us, you know, cross our finger
02:44using AI. Thank you, Alex. Alex Palufo, I'm an associate partner at Quantum Black, which is McKinsey
02:55AI division. I lead very technical teams of scientists, data scientists, data engineers, and what we do,
03:05especially in the scientific AI practice of Quantum Black is we build drug discovery platform with our
03:11clients, clinical trial optimization platforms, target ID platforms, a lot of work around AI apply
03:18in research. So I'll talk a bit about what I've seen a bit around the world in applying AI in
03:24drug
03:25discovery. Great, thank you. Martina. Hello, everyone. My name is Martina Stella. I'm a head of partnerships and
03:32senior researchers in Algorithmic. Algorithmic is a startup that develops quantum software to make use
03:41of current quantum technology and with an eye with the next development in quantum computers. And we spend
03:50time investigating the best applications in the life sciences where we think current quantum technology
03:58can already give some insights, some advantage. And we like to also combine this with AI to find where
04:08AI and quantum technology can be complementary and supporting each other. Fantastic. Thank you all.
04:15So I think it might be useful for everyone just to kind of level set and have a conversation initially
04:22about, you know, where we are. So maybe we're going to, I'd like to ask all of you this question,
04:27but maybe Akilah, I can start with you. How do you see the state of AI and drug discovery right
04:32now?
04:33What do we see? Yeah, so great question. And, you know, it's incredible to hear the work across the
04:40panelists. As you can see, there's tremendous progress across AI and quantum in drug discovery and
04:46biotech. And in particular in our space, when we're using property prediction tools to guide our work,
04:54I just want to make one clear distinction. A lot of times when we hear AI for anything,
04:59we think of LLMs, large language models. And in the case of AI for drug discovery,
05:05the predominant approach is typically domain-specific tools. So what that means is that instead of training
05:11on language, we're training on molecular structures and based on that training, which is actually
05:17conducted in a wet lab. So this is a truly human and machine intelligence paired and augmenting one
05:24another. Based on that wet lab data, these graph neural networks learn molecular structures that are
05:32best able to, in our case, kill particular bacteria. And these models then can predict for,
05:39in our case, antibiotic properties or can actually design these molecules from scratch.
05:45The reason I wanted to make that distinction is that these tools are highly specialized and they
05:51build off of the decades of experience of our microbiologists and chemists. And I think it's a
05:57fantastic example of where these approaches, where you're taking sort of analog human approaches,
06:03pairing it with machine intelligence, really the sum is starting to be greater than the parts.
06:08So I think to your question, where are we in the trajectory of this technology?
06:12I would say in the last year, at least at our company, we've started to really see
06:18that when you combine the human and machine intelligence, we're getting higher rates of
06:23efficacy. For example, in our last AI run, 16 times the potency of our molecules against particularly
06:29difficult to treat bacteria and one-tenth the toxicity. So I think we're right at that inflection
06:35point where, you know, you can start to see the combination of that expertise being built into
06:41our models, our models starting to surpass our original capabilities.
06:45Thank you, Colta. Are you seeing the sum being greater than the parts now, as Akira observed?
06:50Yeah, I think definitely I do agree. And we see that even in Sanofi, if we look to the molecules,
06:58there are two types of molecules, the small molecules and the big molecules. Today, 70% of
07:03the small molecules are enabled by AI. That's a fact. So yes, on the other side, the big molecules,
07:09it's more complicated because it's biology, which is not easy to predict or to build or even to
07:15understand. We are not yet on the understanding of the topic. But we see that we do have opportunities.
07:22And it's become a fact. And it's become also, I will say, a nice friend because in that domain,
07:30very regulated domain, it was like a devil, the AI. And now we realize that, yes, it's a nice friend
07:38that could enable us, that could help us to find scientific pipelines, molecules, and to bring it to
07:47to the market. And we see it even on the side of once you do have your molecule, how you
07:54could bring
07:55it to the regulator with what we call the clinical study report, which usually takes six months to do
08:02it. And now with AI, it's really some weeks, days that you could do that. The same with once you
08:09have
08:10it and you want to manufacture. We had a lot of papers on the product quality report that also
08:17takes more than month and month. And we see all this accumulation of times. At the end, it becomes
08:24the 15 years that I said at the beginning.
08:27So we realize that there is an opportunity. And we have to use it with guidelines. Yes, the data is
08:36important. And even on that, we see that the AI can work on the AI ready data. Like in our
08:43company,
08:44we started like four years ago. And we were like at speed, like a small speed. After a year,
08:54we multiplied that speed of making our data AI ready by four. And we were really happy. Oh,
09:00by four. Then on 2025, we multiplied by 14. And I was so proud we are there. Then the era
09:08of the agent,
09:09it moves us to 30, multiplied by 30. So honestly, it's moving in a high speed. And what's important,
09:18it's yes, to use this high speed on the clinical side, on the manufacturing, on the commercial,
09:24again, for us to bring this molecule on the market.
09:29Thank you. Alex, you're working with a lot of different companies. You have like
09:33an overview. What are you seeing in the market? Yeah, absolutely. I've been in this space for
09:4010 years. And I think the last three, four years have been really transformative,
09:45especially around what you said around physical AI. I think, you know, AlphaFold came out,
09:52that was quite a breakthrough, Nobel Prize. And now there's an explosion of models around
10:01understanding chemistry, understanding antibodies, understanding genomes. So really going beyond
10:07the large language models, and going into what Jensen calls the physical AI, what others call
10:13scientific AI. I think everybody's talking about it. And so, you know, every single of my clients are
10:21actually doing work in this space already, are building teams in this space. I think we're already
10:26past that stage. Now, the big questions are around, how do you interface this with the lab?
10:33How do you automate the lab? How do you do more experiments? There's been a very interesting shift,
10:41which is, you know, when I started doing this work and started working in computational biology,
10:47computational biology and AI in general was very much of a service desk.
10:51for the ones who kind of remember this era, the biologists and the experimentalists would come
10:55to you and say, hey, can you analyze my data? And then find something in there. And now we flipped
11:01it around on its head, which is, we're actually building models. And indeed, in a few weeks,
11:06you can come up with a hypothesis of new molecules that are really new and exciting. But then you want
11:12to test them, you want to ground them in reality, you want the model to learn. And so that's becoming
11:17the bottleneck now. So I think that's what everybody's thinking about. Of course, also
11:21quantum will talk about it. But just to say that what I see is that AI drug discovery is not
11:28a
11:29question anymore. Do we do AI drug discovery? Do we not do AI drug discovery? It's, it's, it's a fact,
11:34it's table stakes. Everybody's building AI drug discovery team. I think there is a range of maturity,
11:40of course, we'll talk about this. I think molecular discovery is quite ahead. Target ID, we have great,
11:48great work there. I think on the clinical side, clinical data, we have a topic of safety here today.
11:56It's been a bit more slow, but now they're catching up. And so a lot of organization as well, they're
12:02trying to optimize their trial with AI, doing, for instance, causal inference, machine learning,
12:06and other things. So we really see the spread of AI across the entire value chain, trying to really
12:14put it everywhere and then connect the operational topic of writing documents, etc, which is also a
12:19massive acceleration with more of the physical models and the lab. I think that's a bit your,
12:23the tripod. So, so yeah, quite an exciting, super exciting time. But there's still a lot to do. I
12:30always say drug discovery is one of the hardest problem on earth. I think it's harder than sending
12:35rockets to Mars or, you know, beating AGI. There are diseases today, like ALS, Alzheimer, for which
12:43there is nearly no cure. And I find that always crazy as quite an advanced society. So I think we
12:49also need to remain humble and realize that there's still a lot to do. Great. Thank you. Martina,
12:55from the perspective of life sciences, what do you say? So actually linking to what you're saying,
13:01um, so with, with AI, we are at a kind of a mature stage when it's about the models that
13:08have been
13:08developing. I myself am a computational chemist. I've been working with chemistry simulations and AI for
13:15many years. And I remember when these things were starting to being developed and now we reached a very
13:21good moment for the maturity of the models. The infrastructures are building up. They are not
13:27still yet there, but we are getting there. And the pipelines are starting to, uh, arise. And this is
13:34very good because this is the direction we need in order to speed up, um, drug discovery. And this is
13:40maturing. However, there's one thing that's, uh, been remaining the same, and that's the aspect I'm
13:46interested in, which is the data. So we are building, uh, pipeline models and idea and, um, projection, uh,
13:55on data that we have been, um, producing for a long time with, uh, some techniques, with some
14:02methods. I can speak from theoretical point of view, uh, but there are experimental techniques
14:08and so forth and so on. And from the experiment, from the theoretical point of view, I know that there's
14:12room for improvement for life sciences and drug discovery, which is basically a problem that's
14:18evolving over and over again, which means for instance, for the neurodegenerative diseases,
14:23we still don't know how to simulate them accurately. So this is where the problem becomes interesting
14:28for me because it means that there's still room for, uh, looking for better fundamental methods
14:35to describe hard problems. And this is also where quantum kind of sneaks in, in the way of promising
14:43an accurate description of hard problems in some future. But this is the direction I, I like to think
14:51about in terms of life sciences. Definitely excited to get into quantum, um, in a little bit,
14:55but there was a lot of nodding on the panel. I noticed when you talked about sort of data being
14:59the bottleneck. Does anyone else like to, you know? Yeah, absolutely. Our, the performance of our models
15:05is linear with the quality and quantity of our data. And I think this is a unique life sciences problem.
15:12So you can't just kind of boil a bigger ocean across the internet and get, get more data. It needs
15:18to be
15:18generated to some degree for the problem that you're trying to solve and the model you're trying
15:22to build. Um, the data often is very experimentally generated. And I thought that was a great point
15:28that it needs to sort of be even more granular to certain ground truths, right? In science and almost
15:35start to think about it in a different way. For example, in antibiotics, we have a lot of binarized data.
15:40And so that, while that's a really good threshold for decision making, does it kill the bug or not
15:46at an 80% threshold? What we're finding with our models is actually an area where we could really
15:52grow in terms of possibility where the model can start to surpass our human intelligence
15:56is if we had more of these gradation based models. So taking, uh, the, the concentration across time
16:05in a petri dish, instead of just like a binarized outcome, we would actually have far better models.
16:10So that's a uniquely AI problem because we can take much more data and feed it into, you know,
16:16these really complex models. So I think also thinking about, do we just take the same old
16:21assays and, and feed them into our models? Or do we fundamentally get back to a new ground truth?
16:26The last thing I'll say on data, which I completely agree, it's so incredibly important. I think it's
16:31also very siloed in our space. A lot of that is historic, competitive, understandably, you know,
16:38commercially, uh, commercial competition is really driven by your molecules. You know,
16:43your gold in pharma is really the compounds that you develop, um, and the IP you build around them.
16:49But I think now we're in a new era where the data that we're generating is two degrees of
16:53separation from what is patentable, yet the culture hasn't quite caught up.
16:59So even if it's possible actually now to share our data and create these vast data lakes,
17:05especially for commoditized assays like, you know, liver, kidney toxicity, we're actually not
17:11conducting that kind of sharing and innovative sort of federated learning because I think there's
17:16still the sense of, well, what if someone could derive a compound structure from this? And,
17:21you know, that, again, that would reduce our competition. Please.
17:25So, yeah, I think everyone here in the room agree that your AI model is good as your data is.
17:33So it's garbage in, garbage out. So, uh, I can tell you that at the beginning of five years ago,
17:37I create a, I would say a frustration where I said, if there is no AI-ready data, there is
17:42no AI.
17:43And they, they see, of course, competitor, but also our other industries showcasing a lot of AI
17:49use cases, product, MVPs. And I said, there is no, yeah, you will create your AI model, but there is
17:55no value behind that. So the starting point for me was like, we need to have a data foundation
18:03in every verticals, every domain, R and D, manufacturing, supply, and, and, and commercial
18:09and corporate. That was for me, the foundational things. And, and again, on the top of that,
18:15that you can use AI models anywhere, every AI models, there are a lot of them and it will work.
18:21But it's not only about the data as a content, but it's also around the governance behind that.
18:29Because it needs to have a clear structures in terms of accountability, who is the owner,
18:34the data, who is the stewardship. And it's not the data team. The data team is an enabler,
18:39but the business that for me, the main point was like, you need to understand your data because
18:45I am not expert on a domain. I am a data AI expert. So we built what we call internally
18:52the AI
18:53foundry, which is a combination of tools in a unified experience where you could bring your data.
19:02And with the tool using AI agent, you, we can help you to make it AI ready, but with your
19:09help,
19:09it's together. And I'm not the one putting the bar. What does it mean AI ready data? It's 100%
19:16quality.
19:17It's 70% is 20%. I don't know. That's the business that need to set the bar of when they
19:24assess that.
19:24Yes, it's good enough to move forward. So that's first. So then I asked to have a clear governance for
19:33every global function from R and D to healthcare inside. I asked to have an excom member accountable
19:41of the AI readiness in the domain. And we do have a follow up on where we are in the
19:48cursor.
19:49It's moving from zero to 2% and so on till they set the bar. Yes, we are good enough.
19:54Because if you don't put a common accountability, it will not move. Yes, the machine is there.
20:03The technical tools are there, but it's not enough. And you need, of course, to involve from the
20:09beginning, cybersecurity team, legal team, because again, we are a very regulated domain. And it depends
20:19on which area and country you are operating. US, it's not the same constraint as Europe. And even in
20:27Europe, there are some difference between countries. And then if you move to Asia, and again, for us,
20:32the patient, it's a citizen of the world. So it needs to be accurate in a way that this treatment
20:38is for everyone from the beginning, from the first of the value chain. So for me, the data, it's yes,
20:44the content with tools with AI to make it ready to be explored and highlighted by the AI models. But
20:53also, it's around the governance that we have in a company to ensure that we do have the same
21:00framework, the same governance, the same way. Because in the value chain, when you finish your
21:07pipeline on the R&D side, you move it to manufacturing. So you need to have a quality check and
21:13standards.
21:14And these standards needs to be a common one. So that's, for me, it was, it's really the important
21:20one. It's to set common standards on the whole value chain. And then on the top, yes, the technology,
21:27the skills around that to make it available in a way that being flexible. Because if you look at that,
21:34we started by Gen AI three years ago, and we moved quickly on AI agents. I will not change again
21:39my framework because there is a new AI emerging technology. That's what we need really to keep in
21:44mind. And we will never achieve 100% because there are a lot of movement. Like I can give you
21:50the
21:50example when we started. So it was 90% of the data we had structured data. And now 90%
21:57of the data are
21:58unstructured. So there are some adaptation to make. But if your architecture is not flexible enough,
22:06if you don't have the adoption, the governance, the accountability between the owner, the citizen,
22:11I will say, of the data and the stewardship, it will not work. Alex, is there an argument here for
22:16some kind of form of data sharing? I mean, I know the industry is built on sort of like a
22:21best of
22:21advantage and an IP. But, you know, is there is there a case maybe of some kind of data commons?
22:27Yeah, absolutely. So I would say, really, I totally agree with what you said. And I think
22:35AI strategy is becoming data strategy. It's really the AI, I always say, you can get a bunch of bright
22:45people and build an AI engine, you can get, you know, with some of our clients, we use sometimes models
22:51that have been built by academics that are great. But if you don't have any data to power it,
22:55you're stuck, right? And so indeed, for instance, a very good example of where we need more data
23:02sharing is antibody discovery. There, there's been a lot of breakthroughs, you know, with alpha
23:08fall and other things. But these models rely on on some hypothesis around how proteins evolve.
23:15And antibodies are not the product, I mean, the ones at least that we designed these days are not
23:19the product of evolution. And so a lot of these algorithms, they don't work very well.
23:23And so with some of some of my team recently, we were working on this problem. And with our client,
23:28we realized that there isn't a whole lot of data sharing, our client had some data, another client
23:34had some data, but there isn't a whole lot of sharing between people on this type of data. And
23:40especially, so there are some data commons, but something that's missing, I find in the space,
23:46and that's biasing also a lot of these algorithms is there is not a whole lot of negative data,
23:51because there is not a whole lot of incentives to share failed experiments. But that's actually
23:55essential. I had a client I was working with, we were doing molecular discovery actually outside of
24:01life science. And then they were very excited. They said, Look, we have a fantastic data set,
24:05you're going to discover a new molecule. And so you know, they, they brought it to my team. And
24:09then one of my one of my colleague data scientist calls me and he says, Alex, it's 98% success,
24:15there's only 2% failure in this data set, that's going to totally bias the algorithm, how are we going
24:20to deal with this, right? So we had to come up with an entire strategy. And we actually had to
24:25spend time
24:25explaining our client that we needed this, they didn't get it, they were like, No, but we want to,
24:29you know, they didn't necessarily get high exactly the details of the AI work. But it was funny,
24:34they were saying, we want the algorithm to only learn from the success, where we were saying,
24:38yeah, but imagine a child, a child also needs to learn from failures, right? And so you can see
24:43already here, like a topic around data generation. And then to what you were saying is, in large
24:49organizations that have that kind of data, very often, it's stuck behind silos, the data doesn't flow.
24:56Another thing is, now I think the data is starting to reasonably flow, but people are excited about
25:03updating the model. So I have some clients that say, okay, you built this model for me to predict,
25:09for instance, clinical trial success. Is it going to update itself when I get my clinical trial results?
25:14When my competitor gets clinical trial results? Is it going to keep updating itself? How do we put that
25:19in place? That's very complicated. I think the people that are architecting the systems are as
25:25important as the people that are building the AI. So I think having also global data sharing resources,
25:33some governments have done a lot of work around biobanks, for instance. That was a huge breakthrough,
25:38for instance, in target ID and in multi-omics data discovery. But there is still a lot to do.
25:46I think there is still a lot to do. But the last thing I'll say is that sometimes, in some
25:51case,
25:52it's also not necessarily about adding more data, but adding better quality data or different types
25:58of data. Recently, I had one of my team who was working on drug discovery, and we were using
26:04cell expression data, so some more gene expression, etc. And we kept throwing more data at the model,
26:11and actually it plateaued in terms of performance. And we realized we need to generate more diverse
26:16experiments, learn from more diverse population to actually also have a better breakthrough. And then
26:22we saw that. We actually traded, instead of multiplying our data set by 10, we invested the
26:28resources in doing a certain set of experiments that were a bit more expensive, but that actually
26:33translated into a better performance. So you can see through all of these different dimensions how the
26:39data strategy and the data generation strategy is going to become really the key for these models.
26:47Great, thank you. I just want to move on now, Martina, if you could talk a little bit about quantum.
26:51Sorry, Akita, do you want to jump in? Oh, I just wanted to make one dovetail on your comments. I
26:56totally
26:56agree. I think one exciting area is that the active learning potential for, you know, now our models are
27:03starting to tell us what part of the data lake they need to see to, and it's very precise. So
27:09we're
27:09not kind of over curating data. So I think that's another bright area.
27:12Great. Thank you. So just wondering, Martina, if you could just, we could move the conversation on
27:17now for touch on quantum sort of, where does that fit into this conversation? Like, is it still fairly
27:23distant or are we beginning to see sort of, you know, promise? So I would say, I would say in
27:30general,
27:30I don't like to make any, um, do me in claims, uh, quantum computing and quantum softwares, uh,
27:37are, are being developed and we are using them and investigating them and extending them. Do I see
27:43them incorporated in a big pharma pipelines in the next two years? Absolutely not. Is it worth it,
27:51consider them relevant as potential tools that can be, um, coupled with, um, that can help, uh, accelerate
28:01some parts of existing pipeline or would be worth it to think about pipelines that can in the future
28:09incorporate quantum in a sort of seamless way? I think that would be the smart thing to do. Having
28:15saying that, uh, you mentioned active learning and I think this is a very, um, interesting example. You
28:21have been hearing about pipelines. Uh, what we mean, it's usually that machine learning, um, can be
28:27identified as a number of steps where you have a goal, for instance, finding a drug with a special
28:34property and you built a number of steps and these steps made up the pipeline. One of the steps is
28:40the
28:41generation of the training data. And what we have been discussing here is, um, how important is the
28:46quality of the training data? And what I was saying is that, of course, quantum computing has the promise
28:52of giving you more accurate training data because the nature of quantum computer is such that it describes
29:00the molecules more accurately. But I also want to say that current standard computational methods,
29:07they have been developed for decades and they are very good. They are very efficient on standard
29:13computers. So we are not doomed in the sense that we are stuck. But what I would do is to
29:19keep pushing
29:20with our standard computational methods to improve our data set. But I would build pipeline that are ready
29:27to integrate the data set instead of using the classical standard quantum chemistry method, some sort of,
29:37help or, you know, improvement that can come from quantum computer.
29:43So the very first application I would see is quantum computing and AI sort of coming together,
29:49reinforcing each other in the parts where they are kind of lacking. So AI is lacking with the data, quantum
29:56computing is
29:57lacking with the speed at the moment. So we can help maybe in the near future to give more accurate
30:03data for some of the molecules.
30:05So active learning helps with identifying subsets and then those subsets can be sort of
30:14validated or it can help to build better training set if we put some highly accurate quantum computer
30:22calculation in it. So this is the way I see it in a fairly near future. This is something that
30:28can happen because
30:30calculations can be done. It's just not on the very large vast scale, high throughput and speed that big pharma
30:38usually do it at the moment, which is, of course, will take the time that it takes.
30:43Okay, so Akhil, I'd love to ask you a little bit about, you know, what you're doing at Fairbio in
30:47terms of
30:48the gap between AI designing molecules and patients actually being able to take them. So I'd love to
30:55know how much of your focus is on bridging that gap.
30:59Yeah, great question. And, you know, what animates our work at Fairbio is two clocks, really. One is the
31:08time to develop a new drug in any therapeutic area is 13 years on, you know, in a good scenario
31:13versus the
31:15time it takes for an antibiotic to develop resistance is less than six years. So we're
31:19constantly losing this race against resistance with antibiotics. We started Fairbio before
31:26this revolution of AI. So we started back in the beginning of 2021. And it was because AI could actually
31:33accelerate and really target the issue, which is that we need novel mechanisms of action. So we were able to
31:40train our models and we're talking a lot about this on training data that the model doesn't
31:45recapitulate. You don't want your model to just take what it's learning and just feed it right back
31:50out. We train our models to learn on a certain chemical space and then identify related but novel
31:57chemical structures. So that novelty is really built into how we train our models. The novelty is the core
32:03of the issue with antibiotics. So it's a really nice solution fitting the problem. So in terms of where we
32:09are, you know, at Fairbio, we've moved from predictive to generative AI. And I think this is a really
32:14important wave of technological evolution in AI for drug discovery. This idea that we're hearing a lot,
32:22you know, from some of the biggest players in the space like isomorphic and in-sitro and silico
32:28is bespoke drug design. And I think right now with generative AI, I would expect if we had this conference
32:35in maybe 18 months, the trajectory will look quite different. One of our chemists recently said,
32:40it's like cooking with a toddler. We have to teach it so much. Yet, once you do start to teach
32:45these
32:46models, we are able to finally build molecules from scratch. So I think as we start to move things
32:53through the pipeline, I think one area of slight concern for me is patience. Patience from the general
32:59public, from investors, funders, from the media, because we're being judged in the pharma space
33:06and the biotech space on the same timelines as many of the consumer products. It's one of the first times
33:11where a technological revolution is hitting all of us at the same time. And so to be judged on those
33:17kind of timelines when we're trying to move from 13 years, you know, in our case down to half would
33:22really enable us to win this race. I think we have to give this more time to fully explore the
33:29range of
33:29technologies. Again, if generative AI is not designing molecules with the push of a button today,
33:34it could in a year, but we have to fuel even inefficiently, I think. I think cost, we can't
33:41let cost be the key driver right now. Speed, novelty, possibility. I think those are really
33:46the metrics that at least we're trying to track. Of course, cost being good stewards of funding,
33:51but I think more so those other bigger metrics of really thinking about this as a historic moment
33:57in biotech. Thank you. So, of course, I'd love to ask you like to reflect on what Martina was saying
34:03about quantum and really how you keep your teams grounded on delivering with current AI, current
34:09data sets or data capabilities, I should say, whilst also maintaining kind of like genuine momentum
34:15on quantum without it becoming a kind of a distraction for your team. So, of course,
34:20we need to continue to deliver because a lot of things are said and the main element shared it,
34:29it's the speed. So, today we have this capability and ability to use AI which wasn't the case several
34:38years ago because it's not the novelty on the models itself, it's the waves how the models are operating
34:43in a high speed to bring the value. So, of course, we need to continue to deliver and to accelerate
34:49what we had in the roadmap scheduled maybe for five years in a year or in a month. That's once,
34:55but yes,
34:56we need to keep the innovation and to explore and that's why the state of the art we need to
35:01be in but
35:02we need to be ahead also to find new things and quantum is another opportunity and I will compare it
35:11to AI maybe in five years or seven years ago which was, yeah, it's great to do it because we
35:17are not
35:18all in because it's still opportunistic, the mindset wasn't there, the infrastructure wasn't fully there,
35:25it's the same for quantum. So, as you said, there are models that could fit others not yet but the
35:33the important thing that I can I see with quantum it's really will bring us in another world comparing
35:41to AI. We still have a lot of what we call the NP-hard problem, basically the mathematical problem
35:48that we couldn't solve in a deterministic way and in a time acceptable way. With quantum we could do that.
35:56We are waiting for the machine but we can't wait till the machine is available and then let's start it.
36:03So, we need to continue to build on and to keep like the acceleration with I will say 70%
36:09on
36:10acceleration delivery and 30% on keeping ourselves on the state of the art because we need to remember
36:15ourselves when we are testing we are not deploying. Thank you. Martina, do you want to come back?
36:20Yeah, I wanted to add two things. I think one thing is really, I resonate with it, I think it's
36:27really
36:27interesting. I don't know if you have heard a lot about saying quantum advantage, the word advantage
36:33associated with quantum and I'd like to sort of twist the angle a bit and I think what would be
36:41the real advantage in this moment in history is just to keep an eye on what's happening and understanding
36:48what kinds of problems might benefit from, let's say, near-term quantum computing and taking the time
36:55to identify this case study. This is not a trivial problem. There are problems that are difficult to
37:02solve. They are not in your face in a sense. You have to look for it because for some other
37:08problems we
37:09have many answers and they are very good and we should not waste time with answers that we know will
37:15work.
37:16But the advantage is we now have a buffer time kind of thing that gives us the chance to look
37:23into
37:23our case studies and understand where are we actually really hitting a wall and if some alternative
37:31technology can help us in this. And this is one thing that I want to say. The second thing that
37:38I
37:38wanted to say is that I often think about this parallel about where AI was and years ago and where
37:46quantum is today and I arrived at the conclusion that we can make this comparison but we have to be
37:53a
37:53little bit careful because I think on some levels we can do it. I think quantum is very interesting
38:01scientifically and it sounds scientifically. We are studying it and AI is the same. Industries are
38:08sort of starting to be interested in it. That's true. It was the same with AI as well. But it's
38:13not the
38:14same. It's the technology. The technology is different. When AI was developing, the technology was
38:20going way faster. HPC was growing like crazy and so we could see how that could fit how the technology
38:26was
38:27growing. With quantum we don't know. We know a little. We know how quantum computers have been
38:34they have been building quantum computers and we are using it but I don't think we can make this
38:39final projection so easily. This last part is a little bit blurred and we don't really know exactly
38:47how it's going to go. I wouldn't say it will go as AI is going right now. This final part
38:53is
38:53unpredictable which is good and bad in a sense. It's good because it means that we can search
38:58still for how we will see this technology to evolve and it's bad because you have to take a leap
39:04in a sense some time to see where it's going. Okay we are so nearly out of time. You guys
39:12have been
39:12great by the way. This has been such a great panel. I'm going to ask you all one final question
39:17which is
39:18if we were at VivaTech in five years time what do you think might have changed and what do you
39:24think
39:24will remain the same? It's like where will we be in five years and if I could start with you
39:29please
39:29sorry Akilah to put you on the spot first. Yeah no of course I think this really will become patient
39:34centered AI for drug discovery and specifically what that means is that you know maybe give one example
39:41in antibiotics. So we have some great therapeutics for MRSA but none of them are actually oral with the
39:47half-lives that get patients out of the hospital. So you're in the ICU oftentimes for for weeks for
39:53something that's very effectively treated. So we need drugs that actually treat the patients in the
39:58settings that they need them that are efficacious for areas of unmet need. So now that we can actually
40:04take those indications and bring it to the forefront of the process, train our models on those particular
40:11characteristics and indications, I think you'll really start to see that it's not just taking
40:16drugs that are coming out of our discovery pipeline and whatever those might be seeing if we can
40:20commercialize them but really taking a targeted approach that fulfills these areas of unmet need in
40:26medicine and gets to better less toxic and more sort of bespoke drugs. So if I were to summarize that,
40:32I think in five years bespoke generative design will be a reality and it'll be based off of some broader
40:40data sharing initiatives across the field because no one company can do this alone, no one non-profit
40:46or university. So I think this will become much more of a hub and spoke kind of model where we're
40:53all
40:53sort of contributing to a bigger part. Thank you. We're out of time so we're gonna have to be very
40:56quick.
40:56Please, Kelton. So I think we will talk again about innovation, not in the same way because I think
41:03I will be embedded everywhere. No fairness around the data sharing because today is still a huge topic
41:11and I hope, I'm not sure that will be the case, but hope the quantum machine will be there
41:18because much more complex model will be fixed and I guess that's a huge bet that we will be more
41:25human
41:26being on the way of how we are innovating and bringing technology to the world.
41:31Yeah, very quickly. I think maybe to your point, patients I think are gonna play a bigger and bigger
41:37role. Now they have access to some AI. There is a guy in Australia who designed a vaccine for cancer
41:44vaccine for his dog with AI. So I think patients are gonna understand more and more indications,
41:51drugs, they're gonna want to be more involved. You know, we see parents starting start-ups to save
41:58their child from diseases, etc. So I think, yeah, people are gonna be more involved and thanks to AI,
42:05I think, so that's gonna be exciting to welcome them as well into the effort.
42:09Great. Thank you. And for quantum, I mean, we are running experiments on quantum computers these
42:16days. It's not that we are not and we are obtaining values and energies and properties that make sense.
42:22Of course, they are not integrated within super fast pipeline. What I hope to see in the next five
42:27years is that perhaps some of the current AI pipelines that, for example, we develop ourselves,
42:34will start incorporating in some ways some quantum data and we will start to see whether or not this
42:43will really benefit the AI pipeline the way I am imagining it to do.
42:50Okay, great. Well, that's a nice optimistic note on which to end. Thank you guys. Thank you so much.
42:55Great panel and thank you for joining us.
Comments