- il y a 2 jours
From Lab to Market: AI-Powered Drug Discovery in Action
Catégorie
🤖
TechnologieTranscription
00:00Hello, good morning, my name is Melissa Heikula, I'm the AI correspondent at the Financial Times and I'm delighted to
00:07be joined on stage today by Colin Murdoch, president of Isomorphic Labs.
00:12Now if you haven't heard of Isomorphic Labs, they're one of the most exciting AI drug discovery companies out there.
00:19But I'll let Colin tell the founding story. How did Isomorphic get started?
00:25Hey, good morning, everyone. Good to see you all here. So yeah, well, great question. So Isomorphic Labs, we're about
00:30three and a half years old now.
00:32And we were really founded after a Nobel Prize winning breakthrough three and a half years ago, something called AlphaFold
00:40that came out of an organization called D-Mind, where I was working at the time.
00:44So AlphaFold, if you've not heard of AlphaFold, is this amazing piece of technology, AI technology, that can really understand
00:51the fundamentals of human life and biology.
00:53So if we start simply, so if we go from our kind of bodies, we get our organs in our
00:59cells and inside our cells are proteins, which really are the engines of our life.
01:04And it used to take years and years of work to understand how these proteins operated.
01:10And with AlphaFold, we took that down to a few seconds.
01:13And that really gave us the inspiration, the first inspiration for thinking about how could AI be applied not just
01:20in biology more generally, but for drug discovery.
01:22So Demester Sabas and I, Sir Demester Sabas, I should say, he got a Nobel Prize and a knighthood.
01:30We sat down and we were discussing, well, what can we do with this technology?
01:34And we realized drug discovery was an amazing area to focus on because, and if you know this, it takes
01:42billions of dollars to produce each drug, takes many, many years.
01:47And only one in 10, one in 10 drugs that you actually design end up making their way through human
01:53trials.
01:54We thought there was a huge amount to go for, and that was the catalyst.
01:57And we put together a pitch deck.
01:59And I still remember being on, this was around kind of COVID time.
02:03So we were on a video conference and we were pitching isomorphic labs to Alphabet.
02:09Alphabet is the parent company of Google.
02:12And within, I could tell, within a couple of minutes, they got it and they were sold.
02:15We finished our pitch anyway.
02:17And we got the money and then we kind of got going.
02:22That's so fascinating.
02:23How did you end up in this field?
02:25I mean, you're at the frontier of this technology.
02:28What's your journey?
02:29What's my background?
02:30So as we were discussing a bit backstage, my brother and sister are doctors, general practitioners.
02:36And there's a general healthcare theme that runs through my family.
02:40I was always very squeamish, meaning I don't like blood and kind of even ill people very much.
02:44And so health was never going to be an area of focus for me.
02:47But I enjoyed computers and I was good at maths.
02:50And so I ended up doing a degree in electronic and software engineering, really getting into computing.
02:55And about, I guess it was just over 10 years ago, I first met Demis.
02:59This is when DeepMind had, it was just before DeepMind was acquired by Google, actually.
03:05And we got to know each other.
03:06And I was very excited by what DeepMind were doing.
03:09So I joined, I joined Demis and DeepMind.
03:12We were about 100 people back then.
03:14Hugely bigger now.
03:17And so I've had the good fortune, actually, to be honest with you, to be at the forefront of AI
03:22for the last 10 years.
03:25And I've been applying AI to fields from a diverse nuclear fusion to kind of wind turbines and Google search.
03:33But when AlphaFold came along, that was my opportunity to close the loop.
03:37And now I can sit around the table at Christmas and my brother and sister can talk about what they're
03:41doing in healthcare.
03:42And I can say, well, do you know what?
03:44Listen to this.
03:45I'm doing something, too.
03:46Amazing.
03:46You made it.
03:47I made it.
03:48I made it in the end.
03:50So tell me about AlphaFold, right?
03:52How does it work?
03:53Explain it to me like I'm five.
03:55Brilliant.
03:55Okay, so as I mentioned, so we're humans.
04:02If you go inside our bodies, we have organs and cells that do a lot of the work.
04:06And then the things that really do a lot of work on ourselves, the little engines of our bodies that
04:10make you and I, you know, the reason we can breathe now, the reason we can live is because of
04:15the work, these amazing little things called the microscopic molecules in our bodies that does this work.
04:22And their 3D structure, a bit like kind of imagine a bit of a Lego that's put together.
04:27They all have these unique three-dimensional structures.
04:29And that is fundamental to the work they can do in your body because they go around like little machines
04:34and their structure really matters.
04:37Understanding their structure is very, very important for a whole range of different things, including drug discovery.
04:42Now to find, you know, how do you understand the structure of something so microscopic?
04:46Well, what you would do is you'd try and x-ray it or you would try and kind of use
04:52other methods which are extreme x-ray crystallography, for example.
04:57Now that requires millions of dollars of equipment and actually often takes the whole length of a PhD to determine
05:04the structure of a protein.
05:05And often it wasn't possible.
05:07So how does AlphaFold come to the rescue?
05:10Well, with AlphaFold, we were able to use all the amazing data that had been produced before from these earlier
05:16methods, some really deep innovative thinking about AI algorithms to come up with an AI model that rather than having
05:24to spend years producing a structure for a protein, we're able to do that with an AI model in a
05:30matter of seconds.
05:31So we can find structures to proteins, not just repeat what's already out there, but we've been able to map
05:38all the 200 million proteins that are known to science, which is quite an incredible feat actually.
05:44And that is the fundamental basis of a lot of amazing research, including isomorphic labs.
05:49So it's almost like some sort of AI-powered microscope that can see right into our bodies and kind of
05:57really understand the deepest components very quickly and the structure of these proteins.
06:01So AI for science, some pretty deep science here.
06:04This isn't Gen AI applied to science, but this is a whole kind of line of AI research in its
06:11own right.
06:12Definitely AI that's actually useful.
06:15AI that is very usable and for me an amazing application.
06:18I don't know what you will think, but AI for drug discovery and helping to solve disease, I think is
06:23one of the best applications there is.
06:25Absolutely.
06:26Now, let's talk about that process.
06:28So how are drugs developed now and how could this technology change that?
06:33So today, and there are some amazing people working in drug discovery, by the way, some really, really smart people.
06:39And we work with many of them, incredible people.
06:42So discovering a drug, if we go back to a protein, a disease is normally caused by a problem in
06:48one of these proteins.
06:48And when you're doing drug discovery, you're trying to create this exquisite molecule that finds its way through your body
06:55and interacts with this protein in some way to enhance or suppress its function and cure the disease.
07:02So the way drug discovery currently works and has worked is you've got these amazing scientists who, based on the
07:07kind of decades of experience they have, they go, well, do you know what?
07:10I think this molecule might look a bit like this.
07:13Let's kind of like design a few of these.
07:15Send them off to the lab.
07:17Test them, you know, three months later, get the results back.
07:20Oh, it didn't really work.
07:22You know, let me try this.
07:24Great, it worked in that way, but now we've got a problem over here.
07:27And if you have this very iterative process, and that goes on for years and years, and sometimes it works
07:32out and sometimes it doesn't.
07:34With the AI tools we're building, AI power drug design engine, you can do all that at digital speed.
07:40So rather than having to have, like, scientists or material scientists think this through themselves, you can search this enormous
07:47space of potential molecular designs at digital speed, at light speed, and then come back to the drug designers.
07:54So we have drug designers at Isomorphic Labs today and say, you know what, we think this is the right
07:58thing.
07:58I'm the AI, and I think this is the direction we should go.
08:01And then the drug designers can overlay their expertise, and that dramatically speeds up the process, dramatically speeds up the
08:08process.
08:09And this is getting better every day, by the way.
08:11We're still a fairly young company, we've made incredible progress, and we still see a lot of potential for uplifting
08:18this further and further.
08:20So the time and cost it takes to get a drug to the market, I see dramatically reducing.
08:25And one other important point, actually, time and cost isn't the only thing.
08:30When I said at the beginning, only one in ten drugs that make it into human clinical trials, which is
08:36a very important step, get through.
08:38One in ten, so there's a 90% failure rate.
08:42What I really hope we do is dramatically lift that as well.
08:45So we're not just getting to that point quicker, but we have a much higher chance of success when we
08:49actually go into human trials.
08:50What does that mean?
08:51It means patients are going to get the drugs they need.
08:54How much quicker will it be?
08:56I think, well, ultimately what I'd love to get to is kind of one-click drug design.
09:00So, you know, you have a disease, you have a protein you want to target, and with one click, the
09:06AI has worked it out for you.
09:07And it goes, look, this is the drug we think you need to take into the clinic.
09:11You have to do some testing before you get there.
09:13But in theory, I don't see why we can't get there.
09:15Now, you guys have a very ambitious goal, which is, you know, no biggie, solve all disease.
09:21How will you do that?
09:22And are some diseases easier to solve than others?
09:25And what do you need to take into account?
09:27So, yeah, exactly right.
09:29It is a very ambitious goal, but I think a very worthy goal as well.
09:32And I think, you know, ultimately one that's very achievable with the power of AI.
09:37So solve all disease with AI.
09:39Now, today, we're building technology which is really agnostic to any disease.
09:44And we can do that.
09:45And this is very, very different.
09:46This is very different than how you would normally approach it because you would normally have to specialize in a
09:50given disease area.
09:51But with AI, we're really modeling things at a very fundamental level, so at the atomistic level, which means that
09:59we're able to tackle pretty much any disease.
10:02However, today, we're focused on ourselves, on our own internal drug design programs, oncology, which is cancer, and immune disorders.
10:12These are areas where I think we can make some really fast progress and bring to human trials some drugs
10:18that we can immediately start having impact.
10:21And then from there, we'll grow into other areas.
10:23And, in fact, we have relationships with two of the leading pharma companies, Eli Letty and Novartis, but we're working
10:28on some other programs where we're helping to advance drug discovery.
10:32So something I want to say, actually, is we're not just building AI.
10:36We're actually applying this now.
10:37I have people in the office right now using the AI tools, working on live drug design programs for patients
10:43that will get these drugs very soon.
10:45Can you tell me a bit more about that partnership?
10:47Because it seems like a really important one to bring this technology from the lab and computer to actual patients.
10:55It's super important.
10:56Pharma companies, you know, super important.
10:58They're really important partners for us.
11:00So we work with them in a couple of ways.
11:03The first way is they come to Minister Colin.
11:05We've got a drug design challenge we've been working on for maybe years, possibly more than a decade.
11:10And, you know, we could really do with some help because we think there's a huge opportunity here.
11:14Can you work with us?
11:16So we say, yes, of course.
11:17And what we do is we recruit our drug design, internal drug design team.
11:22So these are people that have worked in the industry for some time.
11:24And they use our AI tools to work collaboratively with the pharma companies to design the drugs.
11:30And then the pharma companies run the clinical trials.
11:33So they don't take that drug into the clinic.
11:35They'll test it out.
11:37And then they'll do the manufacturing.
11:38And then they'll make sure that drug finds its way to the doctors, the hospitals, and the patients that really
11:42need them.
11:43That's the first model.
11:44The second model is where we have our own conviction.
11:48We think there is a big opportunity here.
11:50We think there's a large, unmet patient need.
11:53And we'll kick off our own drug design programs at our own risk.
11:56And we have these as well in cancer and immune disorders.
11:59So we progress that design.
12:02And when we get to a point where we think it's good enough to go into clinical trials, we do
12:06that as well.
12:07So we start what's called early-stage clinical trials.
12:09This is really just to prove out the drug works in humans.
12:13Once we've got that proof point, then we go to pharma.
12:16And we say, no, you're very skilled at running large-scale clinical trials.
12:20You do the manufacturing.
12:22And then you get that out to patients.
12:23So that's our second model.
12:25So it's very core to our business model working with pharma companies so we can make sure that ultimately these
12:30drugs get through to the patients that need them.
12:33Fantastic.
12:34And can you tell me a bit about, I mean, you're working in this extremely high-stakes field, right, where
12:39things can go really well.
12:40But they can also go really wrong.
12:42How are you thinking about this, like making this, ensuring this technology is safe?
12:46And used responsibly.
12:49The really good news for me is that, of course, people have been designing drugs for many, many years.
12:56And so there's a very rigorous framework, regulatory framework in every country for the steps you have to go through
13:02and the review steps you have to go through every step of the way when you're designing a drug.
13:06So before you go into kind of full-scale human clinical trials, you do a step before where you're testing
13:11on a small group of healthy volunteers typically to make sure the drug isn't toxic or doesn't have any side
13:18effects.
13:19And before you do that, you do some other tests.
13:20And that's all highly regulated, highly monitored, and it has well-established processes which have been developed over the decades.
13:29So that's something that we're able to leverage very effectively.
13:33And, you know, we take a lot of – we take very seriously, of course.
13:37And what would you say are the main challenges facing this?
13:42So in terms of the clinical trials we've just been discussing, I think the main challenge is really that only
13:48one in ten of the drugs that enter clinical trials are successful.
13:53And that is a terrible statistic.
13:55So imagine in whatever industry you're in, I'm not sure what industry you're in, that you produce a product, you
14:00go and actually take it to your customers, and actually only one in ten work.
14:03And so that's a statistic I would love to change because in every case you're trying to create drugs for
14:09patients that really need them.
14:11And that's one area where I think we can have a huge, huge impact with AI because we're able to
14:17– for example, let me give you an example.
14:20One of the problems you encounter is negative side effects.
14:23The drug might do something useful, but then it could cause horrible side effects.
14:27Side effects are normally caused by a drug not just interacting – remember I said drugs interact with the proteins
14:34to kind of help patch them up, but they do something else in your body.
14:38Maybe they kind of interact with something else and they have a negative side effect.
14:42To test for that currently, you typically do these kind of lab-based tests and clinical trials.
14:48But if I'm now doing this on a computer, with the technology we're building, I can test that drug against
14:55every single protein in the human body, the entire proteome, every time I create a design.
15:00So I know when I'm approaching clinical trials, the chances of side effects have been dramatically reduced.
15:07And that is extremely powerful, as you can imagine.
15:09So that's something I'm super excited about as well.
15:13Excellent.
15:14And, you know, these AI models require tons and tons of data.
15:18And with AlphaFold, you had an amazing data set of proteins.
15:22How are you fixing this data problem?
15:24Are there any data sets available in the UK or Europe, for example?
15:29Yeah, so it's a great question.
15:31You know, as you know, if you've worked in AI, the kind of raw ingredients are data, computers, and then
15:36algorithmic developments.
15:37And so data has been a really important part of what we've done.
15:40AlphaFold itself was built on a publicly available data set.
15:45And we've been building on AlphaFold since.
15:48So there are other public data sets out there.
15:50So the UK Biobank is a great example.
15:52This is an incredible data set that UK has collected through the years.
15:56It gives information about kind of genetics and what might happen and follows people and see what diseases they get.
16:03So that's been a very important public data set.
16:05But we're also generating our own data.
16:07So this is data that we generate from what we call wet labs.
16:12So we have, imagine a kind of big laboratory with lots of test tubes.
16:17If you remember back at school, you were kind of mixing different chemicals up.
16:22That's exactly, so we don't have that ourselves, but we have about 20 companies we work with who I haven't
16:29actually been to visit.
16:30But I do imagine like huge kind of chemistry labs like you had at school with lots of test tubes
16:34and bits and pieces collecting data for us.
16:36And so we spend a large amount of time and effort and resources generating this data, which we can then
16:42feed into our models that should improve them.
16:45And that gives us pointers to where the next data set we need is.
16:48That's really cool.
16:49I didn't know that.
16:51You talked a bit about clinical trials.
16:53Now, what stage are you guys at and what can we expect and what's sort of the next step?
16:59So clinical trials, so we were founded about three and a half years ago.
17:04Step one was to build the team.
17:07Step two has been then building the AI drug design engine and then having our drug designers start to design
17:13drugs.
17:13And that's a process that if you're designing the norm, it would take a number of years before you approach
17:19clinical trials.
17:20But I'm very excited because we are now staffing up for getting our drugs into clinical trials.
17:27So that's really on the very close horizon.
17:29That's normally a really big phase shift in an organization like this.
17:32So you've been gone.
17:33You've gone from kind of the lab and doing this work in the kind of virtual lab as we think
17:38about it.
17:39So then starting to work with networks of hospitals across the world to run clinical trials, which, as you can
17:49imagine, is just a very different regime.
17:52And kind of you're beginning to really interact with the outside world.
17:55So that's on the cards.
17:56It's coming very soon.
17:57That'll be in oncology and immunology, our focus areas.
18:01I'm very excited by that.
18:03In terms of the core technology, we're continuing to really push on that.
18:07There are different types of medicines, as you may be aware.
18:10You know, you can take a pill or you can have an injection or you can have an intravenous drip.
18:16All of these kind of work for slightly different diseases in different ways.
18:20So our technology actually is capable of designing drugs across this whole range of different modalities, we call them.
18:29And so part of what we're doing also is expanding.
18:32We started off with oral medicines or small molecules, as we call them.
18:35But we're expanding into these other modalities.
18:38So that's also something that we're pushing on very hard at this point in time.
18:42Maybe one third area.
18:43So we've got clinical trials.
18:44We've got expanding into these other modalities.
18:47A third area is really understanding biology much more deeply.
18:51So there's a lot of diseases out there we know about already.
18:54We know we've got we think we've got a pretty good idea what causes them.
18:59So we can create and design a drug to fix that.
19:02There's actually quite a lot of diseases and quite a lot of human biology where you really just don't understand
19:07very well at all.
19:08Even with alpha fold as it is today.
19:10Like what kind of stuff?
19:12Sorry?
19:12Like what kind of stuff?
19:16So, you know, even some form of cancers, we still don't really understand how they work.
19:21So they may present with the same symptoms, but the mechanism, the kind of molecular mechanisms that are causing them
19:27can be very different.
19:28And that's really important when you're trying to select and design a drug.
19:31You've got to look past it.
19:32We've got to begin to look past the superficial symptoms, understand the mechanisms that are happening within the cell that
19:39are really causing them.
19:40So that's where we're investing in next, a sort of kind of deeper understanding of how a cell works, almost
19:47a virtual cell at some point.
19:49And that's a that's a really big long term, long term investment for us because biology is just hugely complex.
19:56I think we'll be working on that for some time.
19:58But that'll, again, further expand the ranges of disease that I think we can really begin to tackle.
20:04That's cool.
20:05And so concretely, how will that change what we're able to do now?
20:10So what I what I really hope we get to is a place, ultimately, we'd love to solve all disease.
20:14It's our mission.
20:16Step one, of course, is being able to build this drug design engine where I hope we can get to
20:20at one point.
20:21Once we've understood the once we've understood the disease, rather than taking years and years to produce a drug and
20:26then only having a 10 percent chance of success through clinical trials.
20:29We're able to, with a click, design a drug and that's going to clinical trials and that's to be have
20:36a very high chance of succeeding.
20:38And then naturally, what follows from that and the biology investment we're making is a much more kind of personalized
20:43form of medicine because we're all very different.
20:46Our genetics are different.
20:47Our environments are very different.
20:49And what might work for me might not work for you.
20:51So I think the kind of natural progression here is towards a much more personalized style of medicine where we're
20:58all kind of getting the right medicines that we need at the right time and we're healthier.
21:02We're living longer, healthier lives.
21:03And I can say to my brother and sister, there you go.
21:05I did it, too.
21:07And so just to wrap my head around this, is it this idea that you may be able to almost
21:12do a sort of computer simulation of my body and my genes and my cells and then use a computer
21:18to figure out how medicines might work for me?
21:22Yeah, that's right.
21:22That's a really interesting way of thinking about it.
21:24So like some sort of digital twin of you understanding more about you and your particular genetics and your particular
21:31kind of body makeup and create the right sort of medicine that is relevant and tailored to your particular setup.
21:38Because medicines today are designed kind of in a more generalist form and they work in different ways for different
21:44people.
21:44So being able to have that much more kind of personalized header approach is going to have a better outcome
21:49for you.
21:50It'll work better.
21:50It'll have fewer side effects.
21:51You'll get healthier more quickly.
21:54That sounds amazing.
21:55When are we going to get this?
21:57Well, we're making great progress, as I said.
21:59We're already getting medicines coming towards the clinic now.
22:02I don't know how long it's going to take for us to completely crack the whole of biology and do
22:07what we said there.
22:07That's certainly in our mission.
22:09But I think the future is really – I'm very excited about this area, as I said at the beginning.
22:14I've worked in a whole range of different AI application areas.
22:17But I think this is one, for me, that is, you know, just so important for, I think, for our
22:24work, for hopefully all of you.
22:25I think we can all relate to this and for humanity overall.
22:28So in other words, we're working on this as fast as you can.
22:32If you've – you know, I sometimes think if you had a relative or a loved one or someone that
22:37had a disease you were working on,
22:39like how hard would you work, you would work very hard to make sure you're moving that as quickly as
22:43you can.
22:44I much prefer this AI use case to chatbots.
22:49That's fantastic.
22:51Okay.
22:52Now, what advice would you give this audience?
22:54What should they be paying attention to in the next six months?
22:58So I think we are actually in a new era for AI and drug discovery.
23:04AI and drug discovery, you know, there's been a few attempts at this and there's some great companies out there
23:09that have done this.
23:10But the advent of AlphaFold and I think where AI technology has developed today – and I'm not necessarily talking
23:18about Gen AI in general here,
23:20but some of the underlying architectures that have been invented over recent years have really given us a boost.
23:28So I would say if you've been watching this field for a while, I don't know if you have or
23:32not, and you've seen some kind of ups and downs,
23:35I genuinely believe, based on my understanding, that we are, you know, we're in a new dawn and it's very
23:42exciting.
23:43Well, that's a great note to end on.
23:45Thank you so much, Colin, for this fascinating conversation and thank you so much for joining us today.
23:50You're welcome.
23:50Thank you, everyone.
23:51Thank you, everyone.
Commentaires