0:00
In AI, if you don't feel limited by data or you don't feel limited by compute, you're doing something wrong.
Search this video
0:04
When we start using our models and our generative models and our drug design engine, we're making huge amounts of progress in just a matter of months.
0:11
>> Where are you spending most of your resources and time?
0:12
Which which disease areas and what do you think you'll solve first?
0:16
>> You know, what are some of the early wins that we might see?
0:17
How might it progress towards that kind of glorious end state?
0:22
>> How does that change the way we want to live, our society, what jobs we have?
0:26
You know, this is the stuff that's going to be changing over the next 50 years.
0:30
You know, you have to have patient safety first and foremost in mind as you create a new medicine and you want to see if this works.
0:35
Welcome to Giant Ideas, the show where we talk about the ideas that are changing the world and the leaders behind them.
0:40
Today, we've got a very special guest.
0:42
We are talking to Max Yardberg, who is the president of Isomorphic Labs, which is using AI to discover medicines that will hopefully one day solve all disease.
0:53
It's an incredible vision.
0:55
Uh the original financial backer is Alphabet, parent company of Google.
0:59
And the idea and a lot of the team came out of DeepMind here in King's Cross in London where we're talking to Max today.
1:05
So we're going to talk about how quickly AI is going to solve disease.
1:09
Um what is the path to making that happen? What are the obstacles?
1:14
And then talk a little bit about Max's amazing life story and the the challenges and opportunities of leading a company right at the forefront of artificial intelligence.
1:22
So Max, welcome to Giant Ideas. >> Great. Thank you for having me.
1:27
>> I guess let's start with um with the big idea, the giant idea.
1:29
Like you you guys have talked uh various >> small thing of solving all disease. >> Solving all disease. Yeah. >> Yeah. Little one.
1:36
>> The question is how long?
1:36
I mean there's been some amazing statements.
1:38
You know, I think your your colleague Dennis has has talked about getting things done within a decade perhaps.
1:43
Dario for anthropic has been I think even more more kind of bold and said maybe even within 5 years.
1:49
How quickly before you guys are going to solve all disease? >> Yeah. Yeah.
1:54
And you know obviously a mission like solve all disease is a crazy audacious mission.
1:58
Um it's a crazy statement but you know we do say it with a straight face like we do mean it.
2:02
Like that is our north star.
2:05
That's what we're pushing every day towards.
2:06
Um, and I think, you know, fundamentally, like we have a very different technology in hands now that we've never had as a species, which is obviously AI and and we're still at the very beginning of what we can do with AI and deep learning and all of these models.
2:21
Um, and so, you know, what we can see is there's this amazing trajectory and to me this all kind of the first big proof of concept here was of course AlphaFold and seeing that emerge.
2:34
Um but there's this this amazing trajectory of us being able to understand more and more about disease about biology, disease biology, what are the mechanisms of disease and then not just understand that but then how do we actually design you know the medicines that are going to change the those disease states themselves to modulate um our biology.
2:55
And so when you look at that trajectory, yeah I you know I'd really hope 10 20 years we've got the tools in place like we have the understanding that's completely different from where we are today understanding of how disease works, how biology works and then how to engineer medicines and preventative measures to mean that like this we're in a completely different state at that point in time.
3:16
It's obviously really hard to put an exact time frame on it, of course.
3:20
But if you're more generous and you say 30 years, 50 years, if you start to go that further into time, are you, you know, what's your kind of confidence level?
3:28
Are you 100% confident that, you know, within the next 50 years, say humans will eradicate all disease through AI?
3:33
through AI? I'd be severely disappointed if you know we uh like disease doesn't mean something completely different to us at that point in time because you know these really severe late you know diseases that afflict a lot of us you know cancer neurodeenerative
3:47
um you know you know all these like chronic inflammation diseases in immunology like these sort of diseases like they shouldn't exist right we should be able to solve them they're really really hard so if we go 30 years in the future Yeah, I mean I I I really hope that looks very very different. >> And and when you say solve all disease,
4:05
>> And and when you say solve all disease, what what does that kind of mean in practice?
4:08
It means solving cure combining cures for the things that end life early or like how how would you kind of define that more practically? >> Yeah. Yeah.
4:16
And you know this I think this is a this is an interesting conversation because it um the we think a lot about uh you know maximizing health span.
4:25
So, you know, for the years that you're living, like how many of those years are you actually living at your peak potential, you know, or that you think you're really thriving in?
4:35
Um, and of course, you know, very sadly, a lot of people towards the ends of their lives, like they're afflicted for, you know, a long time, 5, 10, even 15, 20 years, um, in a state where, uh, where they don't feel at their at their peak.
4:50
And I think that's because of a lot of these diseases that we, you know, don't have great footholds on.
4:54
we maybe have starting ways to to treat or or to mediate their effects, but I think there's a lot of potential to to change that.
5:03
And that's before you even get to things like, you know, extending your lifespan, longevity and and all of this uh side of things.
5:10
Um and that's also before you get into the um space of uh preventative medicine.
5:16
How do you actually even prevent the onset of these sort of late stage diseases uh in and of themselves?
5:22
And I mean the arrival at the end point of of solving all diseases is just an amazing amazing mission.
5:27
We all hope we get there.
5:29
Like if everything goes to plan in the in the kind of master 10 year plan you guys have internally isomorphic.
5:34
Um uh you know what are some of the early wins that we might see?
5:38
How might it progress towards that kind of glorious end state? >> Yeah. Yeah. Yeah.
5:42
And it's it's such an interesting industry because um you're basically working on these like collection amazing like portfolio frontier science projects.
5:51
Each one of these is trying to create a new medicine.
5:53
We're running many of these in parallel.
5:56
Um, and they uh so so what we can see internally is that what we can see this drug design engine that we're building is really starting to work.
6:06
Like we're able to create these scientific breakthroughs on the ground, you know, at a crazy rate now.
6:10
You know, unlocking progress on some of these uh, you know, drug programs that we've seen that, you know, other people haven't been able to crack before.
6:19
So that's >> you're kind of modeling them in silica or with how how does that work for our for our listeners? Kind of down a bit.
6:25
>> It's a it's a it's a great question.
6:25
So how do you actually make a medicine?
6:28
What what even is a medicine? Right.
6:30
Medicine is a molecule, you know, that could be a small molecule, something you take as a pill that you can actually swallow.
6:35
It can be uh absorbed through the gut.
6:37
Um or it could be a something called an antibbody or a peptide.
6:40
Often these are injectables, so you have to take them in a different way.
6:45
But these are essentially molecules, little molecular machines that go through the body and interact with the, you know, proteins which are the fundamental building blocks of life, the the things that make up our cells, the little machines inside of us.
6:57
Um, and you know, change the way that they work.
7:01
And um, you know, the current state of play I think in in general for making medicines is that this is a very empirical process to find these amazing molecules that are medicines. So what does that mean?
7:16
It means that you have to uh you know come up with a long list of potential ideas.
7:20
You know hundreds of thousands, millions of molecules.
7:23
You put them into a wet lab. You test them out.
7:26
You know you prepet them into different test tubes or plates of cells and you just you know very very fancy trial and error, empirical trial and error. See what works.
7:37
>> And we've got really far doing that.
7:39
We've got some amazing breakthrough medicines that have completely changed our quality of life over the last, you know, 50, 100 years.
7:45
And I still believe it's it's it's just such an empirical process that that's limiting what we're able to actually design and develop and engineer in these in these in these medicines in these molecules.
7:58
And so we see a quite a different process emerging and that's what we are doing internally where we are doing much more design on a computer.
8:09
So we have you know really powerful generative models design agents that are able to come up with the designs of molecules and then instead of just putting you know testing out hundreds of thousands millions in the lab we have these really strong predictive models.
8:26
Think of them as like world models or simulations of the real world.
8:30
Uh alpha fold is an example here which you know means that we can actually just say hey does this molecule work you know when we use our world model rather than go into the real world that means we can do much more design on a computer refine it much further.
8:43
So when we do go to the real world and we go into the lab and we make the molecule and we test it painstakingly there's a much higher probability that it works and that it works the way that we anticipated.
8:54
Uh so you can imagine of course that can speed up the process because you don't have to go back and forth from the real world as much but also I think and most excitingly right now there are you know some problems that are just so hard to solve with that real world empirical process that we can see actually these generative models and these predictive models are able to make progress on and unlock.
9:16
So you know these these are some of the problems where you know some of our big farmer partners for example uh they might have worked on them for like 10 years or 20 years and made a tiny bit of progress um meaningful progress but you know small amount when we start using our models and our generative models and our drug design engine on that we're making huge amounts of progress in just a matter of months.
9:39
So it's been very exciting.
9:42
>> That's a really clear explation I think of how how it works in practice.
9:43
AI for drug design has been around for a while now and I think there's been kind of hype cycles and and before it feels like the last generation has not quite worked out the way people were incredibly excited about.
9:56
Some of those companies have sort of promised a lot but not delivered anything really that much yet.
10:02
Why do you think it's now different?
10:02
I mean clearly AI is a long way forward.
10:06
Is is it just simply that AI is so much further ahead than it was before or is there anything else that's changed?
10:11
Yeah, you you know uh the whole idea of doing more drug discovery on a computer whether it's with AI uh you know more traditional machine learning uh physics simulations that's been around for you know a long time.
10:25
Um kind of the crux is none of those methods worked very well.
10:30
>> Uh you people were using machine learning before using AI before but these were you know what I would call local models.
10:36
So they uh chemical space you know the space of potential molecules that you could design for the space of proteins is is huge like and it's a combinatorally large uh space and design problem.
10:50
Uh and so >> large is probably an understatement.
10:54
>> Large is yeah you know some some of the estimates are like there's 10 to the^ of 60 you know drug-like molecules that could exist.
11:01
You know that's that's a ridiculous number.
11:02
It's too too large to fathom.
11:04
Um the point being it's very big. Yeah.
11:07
>> Um, and >> classic British understatement.
11:12
>> Um, I guess we're in safe company here.
11:14
Um, and the um, and and people were making some progress by training small models around local areas of of of uh, chemistry of of of molecules.
11:26
So, you're working on a drug discovery program.
11:29
You generate a little bit of data.
11:32
You train a model on that and maybe that helps you do the next iteration of this and that would that can improve things by you know a small amount but read like 5% 10%.
11:43
That's not stepch changing our understanding of the space.
11:45
I think that's again where alphafold was the first model that really showed what I call strong generalization goes from a local model to something that generalizes across the whole space.
11:57
So you train alpha fold on, you know, a portion of training data, but then crucially, very crucially for scientific work, you can apply it on completely new stuff that it's never seen before.
12:10
And that looks nothing like the stuff it's been trained on.
12:14
Um, and having access to those sort of models, I think, in biology and chemistry only really started around that alpha fold moment.
12:23
Of course, we've been building on that really strongly.
12:25
You know, we've got half a dozen plus alpha fold like models, level models internally and the the whole focus is on generalization.
12:35
You train it on some data, but then crucially you can apply these models on completely novel parts of science and and biology and chemical space.
12:44
>> And that's the new that's the the really kind of new critical innovation.
12:45
That's the new critical innovation that takes it from yeah, okay, we can make it work and but it doesn't really improve the process by too much to this unlocks stuff that we just never could unlock before.
12:57
>> And is that because the AI is generalizing to like a deeper understanding of biology than we had previously?
13:02
I mean, I'm trying to put my hat on as like a skeptical farmer executive and they might say, well, you know, there's not enough um real world data in the model. Yeah.
13:10
>> Uh or um we don't have a deep enough profound enough understanding of biology for this to actually work in the real world.
13:15
What would you say to that sort of skeptical farmer executive? We've met a few of them. >> Yeah. No, I'm sure.
13:19
>> Yeah. No, I'm sure. And and I mean that is like I think the secret source and the art of like what we've been doing at at isomorphic is you know amazingly and machine learning is still and AI is still magic to me is like we can create
13:32
these models where we can train them on you know a fixed set of data that we know about and we're really clever with our algorithms and how we design our our models and how we design our data distribution and how we use compute and scale that compute. we can actually get
13:44
we can actually get these models to be really predictive, you know, to the point that, you know, a lot of our models are getting to experimental level accuracy.
13:53
That's the benchmark for us.
13:55
We want to replace the real world. >> Yeah.
13:59
And they can get to that accuracy on areas of space that no one's ever seen before.
14:03
And then crucially, it's not just about benchmarks.
14:05
We can actually use this in the lab to design, you know, completely new molecules that interact with uh which recruit like uh uh completely novel biology.
14:16
So, new mechanisms that's no that no one's ever discovered before.
14:20
>> And we haven't just seen that once at this point in time.
14:22
We've seen this, you know, dozens of times now.
14:25
>> And you don't feel limited by the the data that you have access to >> in AI.
14:28
If you don't feel limited by data or you don't feel limited by compute, you're doing something wrong. Yeah.
14:33
>> Um and you know especially in the space of biology in chemistry there aren't these you know you can't just go to the internet and download data here.
14:40
Um there are some you know amazing historical data sets um that have been created but actually we believe that a lot of the data that's going to power this space needs to be generated specifically for the models.
14:57
No one's really thought about generating data for machine learning models >> in biology or chemistry >> synthetic data >> to be honest like real world >> biological and chemistry data. >> How do you do that?
15:07
How how are you working on that to get your hands on that?
15:10
We um you know we have some really really ambitious projects that actually um sort of systematically go and generate this data and we work with some amazing people to do this and um again because we have these models we have our drug design engine internally we know exactly the data that's needed to make them even better than they are today.
15:32
>> I guess the other limiting factor would be the kind of clinical trials stage of all this.
15:36
So again, you know, some would say that you can design everything you want, but at the moment, what really slows progress down on on drug development is these very long clinical trials that just take forever.
15:46
Do do you feel that um that's something that you can work on as well?
15:51
Uh and if not, do you feel like again sort of some kind of frustration that you can design the best things in the world, but you're still stuck by this quite inefficient process that the pharmaceutical industry has of slow clinical trials? >> Yeah.
16:03
No, it's it's a great point. you.
16:06
It would be a sad world if, you know, we're able to design all of these amazing new medicines, but then we're still fundamentally blocked by, you know, just the bandwidth of clinical trials and and how slow they are.
16:14
But they're slow for a really good reason.
16:18
You know, you have to have patient safety first and foremost in mind as you create a new medicine and you want to see if this works.
16:24
Um, and uh, you know, we have to go carefully into patients at this point in time because I don't think we're actually very good at making medicines in general.
16:35
>> You know, 90% of medicines fail >> when they go into the into clinical trials, >> right?
16:40
>> Um, and that's because, >> you know, the way that they thought they would modulate disease biology doesn't turn out to be the right way.
16:48
The molecules that they thought were safe are not safe and they have adverse side effects.
16:54
you know, whether that's um from interacting with other parts of the body or um having other sorts of toxicity and and so therefore you have to go really carefully into patients and that takes and and prove this out systematically and that takes you know a couple of years at the moment.
17:10
Um, now we 100% need to be changing the way that we do clinical trials, but that will only come, I believe, by showing that we can actually make medicines much, much better so that they don't fail 90% of the time. >> Yeah.
17:26
>> You know, if we say that they're going to work in the way that we think they do actually work that way.
17:28
If we design a medicine and we think it's not going to be toxic, it really is not toxic.
17:34
Um so we have a much better handle on you know designing these exquisitly engineered molecules that will lower the failure rate in the clinic and over time mean that we can actually get these drugs to patients a lot faster >> right and is isomeorphic vision then to do the full vertical integration. >> Yes. Yeah.
17:52
You can think of us as like the uh the sort of SpaceX play in in in in sort of drug discovery and development where in making medicines where >> you know the full vertical play um we're you know making the models generating the data designing the molecules taking them through clinical development um and yeah see where we get to >> and will you exclusively go down that route or you you do also partner with farmer incumbents will you just do both?
18:16
Yeah, I think I think doing both is really important because >> the space of disease is really large and uh you need a lot of um you know expertise especially as you go into clinical development on specific disease areas.
18:30
So we focus a lot internally on oncology, immunology, inflammation.
18:37
These are areas where we can build up our bespoke clinical development expertise.
18:41
But there's many other disease areas we want to work on.
18:43
Our AI drug design engine is general in terms of the disease areas it can work on.
18:49
It can work on any disease area.
18:51
So we can then work with you know the best of the farmer world who bring their expertise in you know another therapeutic area and allow us to you know still apply our engine there. >> Yeah.
19:03
And the are the pharmaceutical companies sort of terrified by you, excited by you both? >> I don't know. I don't know. You'll have to ask them.
19:11
I think, you know, >> they're partnering with you. Eli Liy. >> Yeah.
19:14
Well, you know, we we love we partnered with people who are really excited, I think, about this prospect and and have had the >> sort of early conviction that there's going to be something uh in the AI space that that is relevant for them.
19:25
Um, and yeah, I think there's, especially this year in 2026, I think everyone in the biioarma world really understands that, you know, and we've had this conviction for, you know, 5 years, but there's there's no future where we're not designing all medicines with AI.
19:42
It just doesn't exist in in in my mind.
19:45
I can see how well this technology works.
19:48
We're not going to take a step backwards. Yeah. >> From that.
19:51
>> What you've mentioned some of the areas you've worked on, oncology, uh, inflammation.
19:54
where where are you spending most of your resources and time which which disease areas and what do you think you'll solve first?
20:01
>> Those are those are the big ones for us to work on internally.
20:03
So different types of cancer you know different um approaches to solving sort of chronic inflammation uh immunology uh areas and you know we we really focus internally on going after the big things.
20:17
So, uh, you know, solutions for patients that will have profound effect that will really go after some of these really hard problems that maybe people haven't cracked before.
20:30
I think that's our sweet spot at the moment as a company.
20:33
Um, the harder problems, you know, these are the sort of problems that would take 10 years or 20 years normally.
20:40
We're pushing so hard to get those to, you know, patients in the clinic as quickly as possible.
20:46
I feel like, you know, there's at the moment quite a lot of doomerism with with AI and and, you know, suddenly uh it's become the kind of thing to hate in many ways, right?
20:54
Um whether it's people protesting against data centers in Texas or or whether it's this idea that there's a 10% chance of AI killing us all.
21:02
Health and and particularly kind of disease disease prevention feels like the the obviously brilliant use case of AI that will be such a force for good in the world and and you guys are all, you know, incredibly missiondriven.
21:14
the the the flip side to that is is I guess kind of bio resilience and people worry about that about bad actors kind of getting their hands on some of these models.
21:21
How how worried or not should people be about that in general?
21:27
>> Yeah, you know, I I I think there's um we we we really need to make sure that we're using AI in ways that are going to like positively impact humanity.
21:34
That's been a you know, fundamental driving force for me personally for like why we're why we're doing this company.
21:39
Um and you know you have to be really really careful with how these tools are used.
21:47
Um and you know I think there are big implications to the world on the use of AI whether it's in the you know the more general AI or even biofocused AI.
21:58
Um, and I think there's something to be said about actually just making sure that you really have that, you know, security around what you're doing, being really thoughtful about how you're using models, who has access to these things.
22:12
Um, in a company like ours, we develop these models and use them purely internally.
22:17
So, yeah, >> no one has access to that. Yeah.
22:20
>> Um, >> and you know, that really changes the sort of uh security stand, you know, the the risks um from our perspective.
22:26
you guys are working on this for such a force for good.
22:30
force for good. Do you have a kind of visceral sense that wo if this stuff got into the wrong hands it would be very bad for the world and we have to be so careful about that >> at this point in time in this space like I don't have that visceral sense
22:41
>> um and as you said there's always been these sort of technologies over over time >> that can be used in in that way um but it's you know it and and I think part of that is because like we really focus on how on how to like wield these models and train these models purely for our use case. >> Yeah. >> Yeah. >> Right. So, >> sure.
23:02
>> Um, >> so everything we do internally is focused on designing medicines and and understanding biology for for that purpose.
23:09
So, I think that really has focus the application of these models into that into that regime.
23:13
>> I think this is a mission like everyone can get behind uh better lifespans, healthier lives, you know, medicine for all.
23:19
But have you guys ever is there any part of the work in isomeorphic where you talk about some of the unintended consequences if this vision comes to reality and everyone, you know, lives another 100 years healthy?
23:28
God, we'll all be healthier and um enjoy time with our loved ones longer. >> Awful. Yeah. Terrible.
23:35
>> No, I think, you know, there's there's obviously like a a much broader conversation, I think, and and probably people much better qualified than me on like, you know, like where does this whole world go as we're able to um you know, as we have this explosion of intelligence, as we're able to understand so much more about the physical world that we live in, as we're able to live healthier, longer lives altogether.
23:56
Um, how does that change, you know, the way we want to live, our society, what jobs we have?
24:02
Um, you know, this is the stuff that's going to be changing over the next 50 years.
24:08
>> What you're working on, I would say, is both incredibly complicated and difficult for kind of, you know, even the intelligent lay person to understand, but ultimately is just vastly exciting, right?
24:17
Like h give us a sense of of just how profound this can be. >> Yeah. Yeah.
24:23
You know, I I think this is actually Yeah. I'm very biased.
24:24
I think this is like the most exciting mission you can work on.
24:28
Um, you know, this um it it's such a hard problem at so many different technical levels from technology, science, even operationally and how do you put this all together.
24:41
Um, you know, sadly we're not like SpaceX where where you can visually see the rockets going up and and lots of fireworks, but it's that exciting in terms of the types the technology behind this and the breakthroughs that are needed like so many different
24:53
breakthroughs across the full spectrum of science that need to be created inhouse and then come together into this beautiful machine which is a drug design and development engine to bring bring around this new wave of therapeutics. And um you know I think I I personally
25:07
And um you know I think I I personally feel just very lucky to be able to work on this.
25:10
I think it's you know I'm very biased but I think it's going to have you know huge impact on on on humanity over time.
25:16
>> It's really exciting to see the work you guys are doing. >> No fantastic. >> Thanks Max.
25:19
We will be back uh in part two to talk more about you and the way you've actually run the company which I think is also going to be really interesting.
25:26
But thanks so much for sharing your giant idea. >> Brilliant. Thank you guys.