- il y a 2 jours
AI Drug Discovery: Curing with Code
Catégorie
🤖
TechnologieTranscription
00:00C'est OK ?
00:04Hi, everyone.
00:05Great to have you on stage.
00:06First of all, I'm pretty sure everybody knows about NVIDIA,
00:10especially after yesterday's session with Jensen,
00:12which was amazing.
00:14But maybe not everybody knows how much NVIDIA is involved
00:18in the medical field and in medical research,
00:20and that's what we're going to talk about.
00:22And I'm also not so sure that everybody knows about
00:25Isomorphic Labs,
00:26but it's truly a fantastic company.
00:30Its founder and CEO won the Nobel Prize this year
00:34for the research that you are using
00:39and that you are developing for pharmaceutical companies
00:42and for yourself and creating new drugs.
00:44So that's what we're going to talk about.
00:45First of all, Kimberly, can you tell me
00:47what NVIDIA is doing in medical and why your company
00:51that is well known for all sorts of AI and infrastructures
00:55is doing for the medical research?
00:58Yeah, thank you for the question.
01:00Great to see you all.
01:01You just heard from my colleague, Rory Kelleher,
01:05what NVIDIA does.
01:06But let me explain it in a way that you can understand.
01:10We build computers to solve problems that normal computers cannot.
01:16In fact, when I joined the company to start the healthcare practice
01:20some 17 years ago, we were reinventing the company
01:25from what we're known as a graphics gaming company
01:27into what's called accelerated computing company.
01:31And if you think about what accelerated computing is
01:35to solve problems normal computers cannot,
01:38it was really accelerated computing that allowed for AI
01:42to be reborn in the 2012 timeframe of ImageNet
01:47and the breakthrough of deep learning.
01:50And so that is where we started is building these computers.
01:54Now, if you think about solving really hard problems,
01:59where is the hardest problem?
02:01It's in biology.
02:03It's in healthcare.
02:04In fact, look at its humanity.
02:08How could it not be the world's hardest problem?
02:10And so for us to understand what computers we might build
02:15for this very hard problem,
02:16we had to start thinking about healthcare problems
02:20in a way that we could imagine building software tools
02:25and even computing platforms that go inside of genomic sequencers
02:30so that we could really help this industry become a technology industry.
02:36And that's really what we're doing here is building computers
02:39and computing platforms to help the whole industry become a technology industry.
02:44And you heard Rory talk a lot about how that is a very fast transformation.
02:49Yeah.
02:50And we tend to think that these big infrastructures,
02:52these data clouds, these data centers,
02:56we speak of them in a very general way.
02:58Yeah.
02:59But there's a really need to talk, to be close to the research
03:02and to be close to the users for your company.
03:05It helps solve this problem.
03:07They're not that general after all.
03:09They have to be specialized.
03:10That's right.
03:10I mean, if you're trying to solve the hardest problem in the world,
03:13you need to work with the pioneers of that industry,
03:16which is why this is such a great match for us.
03:19We are not drug discoverers.
03:22But we hope that we build computing platforms to enable drug discovery
03:27and do it at new accuracy or new precision or in a much shorter timeframe.
03:33And so, yes, it's very much not general.
03:36It must be domain specific.
03:38AI is not a generalist problem.
03:41Our intelligence is not because we only know what everybody else knows.
03:46What differentiates you and your intelligence is what you study,
03:49what you specialize.
03:50And that's what intelligence is.
03:52And that's why you need to have a healthcare focus
03:55on top of a very capable computing platform.
03:59And, Colin, so maybe we should present Isomorphic Company.
04:02It's a spin-out from a very important company,
04:05probably one of the most important companies in AI, which is DeepMind.
04:09But what is Isomorphic doing especially?
04:12So we are drug discoverers, to Kimberley's point.
04:14And actually, our mission is to solve all disease with the power of AI.
04:19So what does that really mean?
04:20You maybe heard from some of the previous talks that the drug discovery process
04:25is a very, very long, torturous process.
04:28It can take ten years, cost billions of dollars.
04:31Only one in ten drugs actually enter human trials,
04:35make it all the way through.
04:35It's a terrible statistic.
04:37Millions of patients all around the world still suffering.
04:40We want to do something about that.
04:41And as you said, the company was spun out of DeepMind.
04:44It was really inspired by a Nobel Prize-level breakthrough in biology called AlphaFold.
04:51Now, this is a great example of using AI to really speed up drug discovery.
04:55If we start with our bodies, and then we step in one step closer,
04:59we get our organs, and then our cells.
05:01And inside our cells are these little microscopic things called proteins.
05:05And they're really the engine, the engines inside our bodies.
05:09And when they go wrong, that's typically the cause for a disease.
05:12So understanding how proteins work, particularly their three-dimensional structure,
05:17is crucial to the drug discovery process.
05:20And this is where AlphaFold came in.
05:22It used to take years, years of painstaking work
05:25and millions of dollars of equipment to determine the structure of just one protein.
05:30With AlphaFold, we took that down from years to hours,
05:33and in some cases even seconds.
05:35In fact, so fast, we've mapped the structure of all 200 million proteins
05:39can I know to size.
05:40And that, Benoit, was the inspiration for isomorphic labs.
05:44Now, we're doing much more than that now,
05:45but we can really see how AI can dramatically speed up
05:50and allow us to reimagine the drug discovery process.
05:53So AlphaFold is a hard science project.
05:57It's one of the most incredible breakthroughs in biology
06:00because you are able to understand how these proteins interact with each other.
06:05But why choose to make a company out of it?
06:09It's something that also some researchers can have access to.
06:13But what's the specific of having a company doing the medical, the drug design and research then?
06:20Yeah, it's a great question because AlphaFold allows us to kind of study biology in a very broad sense.
06:26And there are versions of AlphaFold out there that are free to use for academia, and that's great.
06:30Because biology is the study of living things, which is a huge field.
06:34But to tackle drug discovery specifically requires a very particular mix of skills,
06:40a very deliberate mission, and a lot of focus.
06:44And we have people from not just an AI background, but as Kimberley mentioned,
06:49we have drug discovery experts, and they work together in a distinctive interdisciplinary culture
06:55driven towards that singular mission.
06:57And I think that's the way you have to build a company like this
07:00when you're trying to really transform a space.
07:03So that's why we built Isomorphic Labs.
07:05We're three and a half years old now.
07:07We're doing absolutely fantastically.
07:09Every week I review our progress and I'm astounded with how much progress we're making.
07:13Super exciting space.
07:15And the aim of the thing is to speed up research.
07:18It's to go to clinical trials faster with better drugs?
07:25Yeah, all of the above.
07:26So it takes many, many years, first of all, and a huge amount of money.
07:29So time and cost are one thing.
07:31But most important is improving the probability that when you design a drug,
07:36and this is terrible.
07:37It's only one in ten drugs that are successful when you put them into humans.
07:40You could even double that, triple that.
07:42You can have a profound impact on the number of diseases you can tackle.
07:46And not just diseases that, kind of, we know about today
07:49and we can maybe tackle today, maybe we've got a drug already.
07:52There are some diseases that the industry, and these are very smart people by the way,
07:57very, very smart people, have been working on for years,
07:59if not more than a decade, and they haven't been able to tackle them.
08:02And that's immensely sad for the patients and their loved ones and their families.
08:06So what AI allows us to do is to find solutions for these diseases that weren't possible before.
08:13A quick analogy.
08:14A drug essentially is something that interacts with that protein that I mentioned a moment ago.
08:19and you're kind of searching through this huge potential space of different kind of combinations of molecules.
08:25More atoms than there are in the universe options.
08:28You just can't do that as a human being. Impossible.
08:30That's where AI comes in and that's where it can make a huge difference.
08:32And that needs an incredible amount of compute powers we all imagine.
08:39Yes. Thank you, Kimberly.
08:40And that's where Kimberly comes in.
08:41So how do you work with these kind of labs?
08:43Do you do some special design of computer clusters for them?
08:50What do you do? How do you work with labs?
08:52Yeah. So if you think about isomorphic,
08:55they are a beautiful expression of what is an AI-first, an AI-native drug discovery company.
09:04And so we also share that future state that we will, like we do with chips today,
09:12when we design a chip, it's completely in silico, completely done in a computer from the beginning,
09:19all of the different circuits working together to get into the board level, the system level,
09:24simulated, emulated in supercomputers that when we actually go tape out that chip, it works perfectly.
09:31Wouldn't you like to recreate that in drugs where you're able to, as Colin described,
09:39essentially explore the entire almost infinite vast space of chemicals and proteins
09:45and the different combinations of them and simulate them for the properties that are necessary
09:50so that when you get to the answer, it's much more likely it's going to work in humans.
09:57It's the same thing that we're trying to create here.
10:00And now that we, over the last, you know, 40, 50 years,
10:05the world of biological platforms and instruments, the whole genome,
10:09the ability to do CRISPR, cryo-electron microscopy are all of these instruments that measure biology
10:18are making way for the data generation in order for us to model biology in a computer
10:25and represent biology in a computer for the first time in history, in mankind.
10:30And that is the opportunity Isomorphic Labs has understood.
10:34That was the critical breakthrough of AlphaFold to help us realize we can start modeling biology,
10:41in this case proteins, and do something, as he described, of going from a genomic sequence
10:48to its 3D structure, which used to take a PhD student their entire PhD term to go and hopefully resolve
10:57one singular protein.
10:59As Colin said, that one powerful Nobel Prize winning model can do now in seconds.
11:06And so we need to, as Colin described, the scales of biology are very complex.
11:13DNA, proteins, RNA, transcriptome, on and on we go into cells, as Rory was describing, virtual cell.
11:21We need to build the methods such that we can represent biology in a way that we can understand it.
11:28build that foundational model and understand it, then we can go off and interrogate it, reason over it,
11:36specialize that model for a certain disease case.
11:39This is the journey that we're on, and I think that it's absolutely, there's no contest, there should be no
11:47contemplation that the future is going to be AI-driven drug discovery.
11:52And how is the world of pharmaceutical companies reacting to that?
11:56I mean, drug discovery is almost as old as medicine, and it has been going through very long processes.
12:03It's improved over the years and the years and the years.
12:05But now, the ability to have these in silico experiments and all these in silico trials completely changed the thing,
12:15and now AI is changing again.
12:17So how do you work with pharmaceutical companies on this subject?
12:21Do they come to you?
12:21How do you organize this research for newcomers?
12:24Yeah, that's for Colin.
12:25But before he answers the question, actually, you say that the drug discovery process has gotten better.
12:31Actually, it's on what they call E-Room's law.
12:35The more they've invested in R&D, the less productive the R&D has become.
12:39This is the true problem statement.
12:41Okay.
12:42So that's factual.
12:43That is written down.
12:44Right.
12:45That's the reverse of Moore's law.
12:46If you reverse the letters, that's how you get E-Room's.
12:49Fun fact.
12:49E-Room's law is the reverse of Moore's law.
12:51Yeah.
12:52So how are we working with pharma companies?
12:55So we are not just building the AI tools, to be clear.
12:58We also have an amazing group of drug designers ourselves.
13:01And we partner with some incredible people like Eli Lilly and Novartis, who's the two leading pharma companies,
13:07where they may come to us and say,
13:08Colin, we've got a drug we would like to design.
13:12We have a particular disease in mind.
13:13Can you help us?
13:14Can you and your team help us out?
13:16And we'll work with them in a very collaborative way to tackle some very, very hard problems.
13:20And I can't share the details of those, but we're making fantastic progress.
13:24And in fact, we just extended our relationship with Novartis.
13:28We also have opportunity, and we do this as well.
13:31We have our own drug design programs.
13:34And we work in the areas of cancer and immunology at the moment.
13:37So trying to create medicines to address cancer and immunology are very relatable, I think, for many of us.
13:44They're investing off our own back to drive forward our own drug design programs.
13:50And when we get those to a certain point, we then speak to the pharma companies,
13:54who are then able to continue the journey, take up those drugs and do broad human trials,
14:00manufacture those and get them out to the patients and doctors that really need them.
14:04So Super, alongside NVIDIA and our compute colleagues, the pharma firms are really, really important partners for us.
14:12You're a very young company.
14:14And how do you estimate time you will need to get your first molecule or your first, yeah, it will
14:21be a molecule in the end,
14:22your first drug on a clinical trial on humans, for example?
14:26Yeah, well, look, we're very close, actually.
14:28We've certainly already got molecules, because that's something we're kind of doing every day.
14:33The AI is proposing different molecules with different dimensions.
14:37And actually, we're going to be announcing very soon that we are kind of getting very close to clinical trials.
14:44We're staffing up to account for that.
14:46So, I mean, watch this space, but within a very short period of time, because we're almost there,
14:52which is really exciting for me, personally.
14:54So, I grew up in a family where I was the old one out.
14:57My brother and sister are doctors.
14:59My parents are from a medical background.
15:01It's very rewarding for me to be able to apply my skills and technology to, you know, really address these
15:07patient needs.
15:08And that's when it becomes very real, when you start putting in people who actually have these diseases.
15:12And we're very, very close.
15:14Yeah, you wanted to add something?
15:16There is evidence.
15:17There are other fantastic companies.
15:20In Silico Medicine is one that has documented their digital platform, essentially generative AI platform, from target discovery to in
15:32-clinic patients.
15:33And they have reduced this down to, I'll call it 18 months to three years.
15:38Something that is $300 million to a couple of million dollars.
15:43So, there is evidence that this is truly happening.
15:47And so, it's wonderful to have incredibly, you know, architecting a company like Isomorphic and these new type of companies
15:56that are kind of called tech bios.
15:58They're tech first, biology second.
16:01They're very platform oriented.
16:03They're building computing platforms that allow them to go through this virtual screening and finding process and tuning process and
16:12super sophisticated combination of, you know, world class Nobel Prize and beyond models that allow them to expedite this process
16:22immensely and go where no human has gone before.
16:25Scientists were kind of locked in their idea of what they could experiment and see in the lab.
16:30This opens their aperture immensely.
16:34You know, not just one order of magnitude.
16:36We're talking many orders of magnitude.
16:39There are 10 to the 160.
16:4110 with 160 zeros after it.
16:45Proteins.
16:45Potential proteins that might be a therapy.
16:48How do you explore that in the lab?
16:51We would never get to it in our lifetime.
16:53And so, that's why we built computers to tackle problems so that the scientists, the brilliant scientists and the brave
17:00ones can solve, you know, medicine, you know, have a drug and solve medicine within a very short period of
17:07time.
17:08And what are the next steps?
17:09I mean, obviously, solving the protein folding problems was an enormous advance.
17:14But there are going to be some breakthroughs and maybe some blocks along the road.
17:20What's the vision for the next step?
17:22Yeah.
17:23I can talk about it from an isomorphic lab perspective.
17:26So, the beginning of a drug design process, you have to understand the protein you're going to drug.
17:32Then you have to design a molecule that interacts with that and what we call binds to it and stays
17:38stuck to it.
17:39You think of these molecules as kind of either inhibiting or enhancing the function of these proteins.
17:43You've got to make sure, actually, when you take a pill, for example, it gets into the right place.
17:47These are all very complex problems.
17:48And you've got to optimize all of those things together.
17:52And then ideally, what you have is an AI that actually does all that in one click.
17:56So, where we're headed, actually, is sort of one-click drug design.
18:00So, once you understand the disease, within one click, you'll be able to have a drug that is ready to
18:05go into the clinic.
18:06That's what we call chemistry, if you like.
18:08That's the drug design.
18:10As you heard in the earlier panel, maybe, and as Kimberly mentioned,
18:12actually understanding disease itself, finding that protein that you want to target with that medicine,
18:19that is a very, very complex process.
18:22Understanding biology and decoding biology, very complex.
18:25So, that is the next area where we're headed.
18:27AlphaFold has already given us a huge lift there.
18:30But we also have investments in that space as well.
18:33So, one-click drug design, if you like, and then decoding biology.
18:38And I think going to a more personalized cure or more niche disease is also something that you will be
18:48working on.
18:48Right, exactly.
18:49This is all possible once you can do what I've described there.
18:53You can have someone, you can decode their genome,
18:56you can understand which might be the appropriate drug for them,
18:59and you can offer that up, and that gives you a much better chance of success
19:02because we're all different.
19:03We all respond very differently to individual drugs, and I think that's a fantastic outcome.
19:07Okay.
19:09And Kimberly, outside of drug discovery, we saw that NVIDIA is on many, many other things.
19:14What's your vision of the future of medicine?
19:17Yeah.
19:17Does it include robots, like we saw some robots yesterday?
19:21It does, yeah.
19:22And there are two types of robots.
19:24There's digital robots, which are AI agents, I think you've heard a lot about.
19:28And I would go so far as when Colin is describing a one-click drug discovery process,
19:35I absolutely could imagine a future like that.
19:37In fact, we built a demo, and we had it at our GPU technology conference,
19:41where a researcher is going to have many agents that go off and do the literature research,
19:48hypothesize about what data that might need to collect,
19:50and go kick off an experiment in the lab, which is completely driven by robots.
19:55Take that data, retrain a model.
19:58After you retrain the model, put it into a virtual screening,
20:01so you can go from a huge amount of ideas down to a small amount of ideas.
20:06And on and on it goes, where you're going to be having scientists, machine learning experts,
20:12working with a fleet of AI agents.
20:16This is how science is going to be.
20:20This isn't just drug discovery.
20:22I'm describing the future of science in that vision there.
20:26And then the healthcare delivery side has every aspect of that great benefit.
20:32In fact, it's very topical in here and now.
20:37If you think about many of the struggles that our healthcare systems globally are suffering,
20:43it is from the too much demand for healthcare, too few healthcare professionals.
20:49This is a global epidemic.
20:52We are in an existential spot with healthcare globally.
20:56Well, what could we do?
20:58We could augment every healthcare professional with AI agents.
21:04Everything from transcribing the conversation in a doctor's office
21:10and formatting it into the expert clinical note that needs to be entered into the electronic health record,
21:16which is a huge burden on our healthcare professionals today,
21:20to allowing for me to call up my doctor, but maybe it's not my doctor,
21:25but that agent could provide me enough information that I could make my next decision.
21:31Agents working in the healthcare system is going to be huge.
21:35And then we're seeing massive opportunity to advance with robotics.
21:39The future of hospitals, the whole hospital is going to be a robot.
21:44It's going to be acting with human actors, delivery robots, surgical robots,
21:52being able to really enhance the operational efficiency on the one hand,
21:57which we need to do for our healthcare professionals,
21:59but to also improve the outcomes because robots are going to be able to do things, again, that humans cannot.
22:05These robotic systems are allowing us to reach nodules in the lungs that a human could just never go,
22:11or be so precise in the brain, which a human just wouldn't have the capacity to do.
22:16And so you're going to see the entire healthcare delivery be reinvented with digital agents,
22:23working with human healthcare professionals as a hired team,
22:27and then you're going to see physical robots operating in the hospital environment,
22:34from humanoids to delivery to just exquisite surgical robots that are going to improve outcomes.
22:42So we're just more than excited about the opportunity to transform both finding of new medicine
22:48and helping people recover from their illness.
22:52Thank you so much, both of you.
22:53We went from the really tiny molecules and the proteins to huge hospitals as robots,
22:58which is quite an amazing journey.
22:59Thank you so much for those interested in AI agents and patients and doctors' relationship.
23:06We have a very interesting panel in the 20 minutes with Dr. Lieb and Google DeepMind also.
23:11Thank you so much. It was a pleasure hosting you.
23:12Thank you.
23:13Thank you so much.
23:16Thank you.
23:16Thank you.
23:20Thank you.
Commentaires