0:00

have set up the company from day one

0:02

to really go after this big ambition. This

0:05

isn't about

0:07

developing therapeutics for a particular

0:09

indication or a particular

0:11

target. It's really thinking

0:13

about how do we create a

0:16

very general drug design engine with

0:18

AI, something that we can apply

0:20

to not just a single target

0:22

or even a single modality, but

0:24

we can apply this again and

0:26

again across any different disease area. And

0:29

that's what we're stepping towards at the moment. Today

0:47

we're excited to welcome

0:49

Max Yatterberg to the show, Chief AI

0:51

Officer of Isomorphic Labs, which launched

0:53

out of deep mind with the

0:55

goal of revolutionizing drug discovery using

0:58

AI. Last summer,

1:00

they released AlphaFold 3, a

1:02

stunning breakthrough that allows us to

1:04

model not just the structure

1:06

of proteins, but of all molecules

1:08

and their interactions with each other. That

1:11

led to Demisosobis winning the Nobel Prize

1:13

in Chemistry last year. Max

1:16

describes their vision for what a holy grail

1:18

model for drug design and what agents for

1:20

science look like. He draws parallels

1:22

to his experiences building AlphaStar and

1:24

capture the flag. and the

1:26

research directions of building agents and games

1:28

more broadly. Specifically,

1:31

with 10 to the power of 60

1:33

possible drug molecule structures, we need to

1:35

build both generative models and agents that

1:37

can learn how to explore it and

1:39

search through the whole potential design space. Max

1:43

also describes his vision for what a GPT -3

1:45

moment for the field might look like. Describing

1:48

it more akin to AlphaGo's famous

1:50

Move 37 when we start to

1:52

see things that exhibit superhuman levels

1:54

of creativity, in AI drug design, and

1:57

that stun even humans ourselves. This

2:00

is one of my favorite episodes yet.

2:02

Enjoy the show. Max,

2:05

thank you so much for joining us today

2:07

here in London. It's a pleasure to be with

2:09

you here. Yeah, it's fantastic. Awesome

2:12

timing, too, with the launch of Alpha Fold 3

2:14

and with Demis winning the Nobel Prize in Chemistry,

2:16

which is a true testament to everything that you

2:18

and your team have done over the last couple

2:20

of years. Yeah, 2024 was

2:22

definitely a busy... for

2:24

us. Lots of big breakthroughs. The

2:26

Nobel Prize was just incredible to see. You

2:29

know, I think amazing recognition for this for

2:31

this seminal piece of work. Yes. Well,

2:33

I'd love to start with talking

2:35

a little bit about your own personal

2:37

story. You've had an incredible career

2:40

in the world of deep learning from

2:42

the very start, authoring many seminal

2:44

papers while at DeepMind, including for Capture

2:46

the Flag and Alpha Star. breakthroughs

2:49

in the world of deep learning. Can

2:51

you walk us through some of the

2:53

key questions that you had in your

2:55

field of research around deep reinforcement learning

2:57

at the time? Yes,

3:00

so at DeepMind, I

3:02

worked on a whole host

3:04

of stuff, early days of computer

3:06

vision and deep generative models,

3:08

but it was really reinforcement learning

3:10

that ended up hooking me

3:13

there. DeepMind was the place in

3:15

the world to be working on reinforcement learning at

3:17

that time. You

3:19

know, really the question in our minds was,

3:22

how can we actually get to a

3:24

point where we could get an

3:26

AI that could go off and do

3:28

any task you wanted it to

3:30

do? And, you know,

3:33

the dominant paradigm at that

3:35

point in time was supervised

3:37

learning. And

3:39

supervised learning is very different from reinforcement

3:41

learning. They're both learning techniques, but

3:43

supervised learning, you need to know what

3:45

the answer to your question is.

3:47

And that's how you train the model.

3:49

So in supervised learning, you give

3:51

an example and then you supply the

3:53

model with the answer to that

3:55

question. Now,

3:57

that can be great if you already

3:59

know everything about the problem that you're

4:01

training this AI to do, this

4:03

neural network to do. But most times

4:05

you don't. Yeah. I mean, there's

4:08

just so many problems in the world where

4:10

we don't know what the answer is. We

4:12

don't know what the solution is. And if you

4:14

think about, you know, I think about how I want

4:17

AI to be applied to the world. Yes,

4:19

it's going to be great to be

4:21

able to apply things where we're already

4:23

good as humans here. But really, you

4:25

know, the big frontier is can we

4:27

start applying AI to places where, you

4:29

know, humans don't know how to do

4:32

this stuff or, you know, there's a

4:34

limit to human performance there. And,

4:36

you know, that's where reinforcement

4:38

learning is. one of the key

4:40

tools and has real promise

4:42

here because in reinforcement learning, you

4:44

don't need to know what

4:46

the answer to the question is.

4:48

You just need to be

4:50

able to say whether the answer

4:52

that the model gave you

4:54

was good or not good. Maybe

4:56

even how good or not

4:58

good. So this opens up a

5:01

completely new field of problems

5:03

to train these models against. And

5:06

so reinforcement learning and

5:08

really starting from what

5:10

was one of the big breakthroughs

5:12

of DeepMind in the early days was

5:14

working on games like Atari. Yes.

5:16

The question was, okay, so how can

5:18

we scale this up from the

5:21

world of Pong and space invaders to

5:23

things that really start to look

5:25

like real problems in the world? And

5:28

so there was an amazing track

5:30

of research as we scaled up these

5:32

methods. Yeah. Did you know that

5:34

Sequoia was the first investor in Atari

5:36

back in the day? Oh, really?

5:39

I didn't know that. That's incredible. Yeah,

5:41

no, those Atari games were great fun

5:43

actually to sort of go back and play

5:45

in the context of, hey, we've got

5:47

an agent and I'm just going to have

5:49

a game of pong on the sides

5:52

as well. There's a

5:54

wonderful wall at Sequoia in our

5:56

office where we have all these

5:58

names of legendary IPOs and M &As

6:00

that have happened. And there's one.

6:03

I think it's called the Pizza

6:05

Company. And I love asking

6:07

folks if they know what that is.

6:09

And it's actually from Chuck E. Cheese's, which

6:11

was an original Sequoy investment at the

6:13

time. Amazing. Amazing. So

6:16

Capture the Flag and Alpha Star

6:18

were incredible breakthroughs at the time.

6:20

Can you share a little bit

6:22

about what exactly those breakthroughs were

6:24

and maybe why you chose those

6:26

specific games? Yeah.

6:28

So, you know, if you think about

6:30

the history of AI, using

6:33

video games. Why do we use

6:35

video games at all? Video

6:37

games are these sort of

6:39

malleable, perfectly encapsulated worlds

6:41

that as researchers and scientists,

6:44

we can manipulate them,

6:47

we can test out different algorithms in them,

6:49

we can set up different situations. So

6:51

the perfect test ground for us to develop

6:53

new algorithms. And

6:56

then you can imagine as a

6:58

RL... is someone who's like thinking about

7:00

how can we get AI to

7:02

be as general as possible. You're always

7:04

thinking, okay, we've cracked Atari, how

7:06

do we get a more complex game? And

7:10

the thing that I was

7:12

personally obsessed with is I

7:14

want these agents to be

7:16

able to zero shot, be

7:18

able to do any task.

7:21

And this is a slightly different paradigm from

7:23

the, from what people were doing at

7:25

the time with training on Atari where Normally

7:28

in reinforcement learning you think about, here's a

7:31

game, now you get to train on it

7:33

and get good at it. And then you

7:35

apply that same algorithm from scratch training on

7:37

different games. I'd

7:39

love a different scenario where instead we

7:41

train an agent and then we can lift

7:43

it and put it on any new

7:45

task and that agent will be able to

7:48

perform well in that task without any

7:50

more training. And

7:52

so to do that, what you're

7:54

really asking for is generalization over task

7:56

space. And

7:58

that means you need lots and

8:00

lots of training tasks. So

8:03

the training data in this

8:05

RL for agents becomes tasks,

8:07

not images, not pieces of

8:09

text, but tasks. And so

8:11

you can imagine you could go

8:13

and sit and take a whole

8:15

game studio and try and hand

8:17

author hundreds of different tasks, lots

8:19

of little mini games in these

8:21

virtual worlds. And

8:23

we did that. We were doing lots of

8:25

that. And then you can

8:27

think, yeah, we can actually go

8:30

further than hand -authoring. We can

8:32

procedurally generate these tasks and games, generating

8:35

worlds and maps

8:37

and different objectives. And

8:39

we did that. But

8:42

you keep running into this

8:44

complexity ceiling that there's only so

8:46

much complexity that you can

8:48

hand -author or you can design

8:50

humanly. But that's where

8:52

multiplayer games come in. Because

8:55

as soon as you go from single player

8:57

to multiplayer, it's not just the agent playing.

8:59

You've got another player in this game. And

9:03

that other player or other

9:05

players can take on many different

9:08

characteristics and many different behaviors. So

9:10

every different player, every different strategy that

9:12

you're up against changes fundamentally the game

9:14

and what the agent is trying to

9:17

do. I go back

9:19

and think, Why are

9:21

people still obsessed with playing chess? Why

9:23

does a professional chess player still keep

9:25

playing chess? It's the same game. But

9:28

it's actually not because you're playing completely different

9:30

opponents day after day and new people into

9:32

the world. So

9:34

the game is continually changing. So

9:37

multiplayer games and multi -agent

9:39

games really encapsulates that huge

9:41

diversity of tasks that you

9:43

might encounter just from other

9:46

players being there. And

9:48

so capture the flag was

9:50

actually one of our first

9:52

forays into How can we use

9:54

multiplayer games to really stretch

9:56

what our reinforcement learning algorithms can

9:58

do to? Really force us

10:00

to think strongly about how we

10:02

can generalize to new tasks

10:04

how we deal with these multi

10:07

-agent dynamics So capture the flag

10:09

was a fantastic breakthrough Really

10:11

showed that we could get to

10:13

human level performance for these

10:15

multiplayer first -person games. Yeah, and

10:17

of course starcraft added on

10:19

a huge amount of complexity and was sort

10:21

of the next frontier that we had

10:23

to go after for this. You were so

10:25

early in this that so many of

10:27

these concepts are very, very relevant today in

10:29

the world of language. How does it

10:31

feel to see some of this work continue

10:33

to be played out? Yes,

10:35

brilliant. It's just

10:37

fantastic, actually. There were so many things

10:39

that we were talking about at time.

10:41

Seven years ago, yeah. You know, 2015,

10:44

16, 17, 18. to

10:48

see all of these core fundamental

10:51

concepts be really useful and really

10:53

applicable today in the world of

10:55

large language models. And

10:57

resulting in performance that we could

10:59

only really dream about at the time.

11:01

That's incredibly satisfying. So

11:04

then in your own words,

11:06

you said that you moved from

11:08

building toys to then finding

11:10

real applications. When did you know

11:12

that you found the right recipe? I

11:17

just love deep learning. I've

11:19

been obsessed with deep learning for

11:21

10, 15 years now. And

11:24

the thing that I love

11:27

about it is that you

11:29

have these underlying core concepts,

11:32

these fundamental building blocks

11:34

that are somehow incredibly

11:36

transferable between different application

11:38

spaces. Yes. So,

11:41

you know, it's the same building

11:43

blocks that we were using in computer

11:45

vision in 2012, as we were

11:47

using in, you know, generative models in

11:49

language, you know, then reinforcement, et

11:51

cetera, et cetera. So

11:54

what I was seeing just

11:56

again and again was this

11:58

ability to take these core

12:01

concepts, these same core concepts,

12:03

take incredible people who understand

12:05

how to you know, they're

12:07

almost like master chefs of

12:09

putting these these concepts together

12:11

in these different building blocks

12:14

together. Take a

12:16

team of incredible people and

12:18

go after, you know, really,

12:20

really challenging problems, you know, problems that you

12:22

go to conferences at the time and you

12:24

talk to leading researchers in the field. They

12:26

say, no, no, no, this is 10 years

12:28

away. And in the back

12:30

of your mind, you know, okay, we

12:32

actually we basically cracked it. Wow. And

12:34

I saw that happen again and again

12:36

and again. You

12:39

know, you take amazing people, amazing

12:41

algorithms, amazing compute on really challenging

12:43

problems and we can find recipes

12:45

now to crack so many problems. And

12:48

so it just got to

12:50

the point where and I've always

12:52

been quite obsessed with the

12:54

application of these methods. I

12:57

want to see this technology

12:59

have, you know, real transformative positive

13:01

impacts in the world. And

13:04

so. we need to start actually going

13:06

after that and the time has been

13:08

right for I think a few years

13:10

now. Well,

13:12

so you've now had a decade

13:14

long relationship working together with one

13:17

of the greatest scientists, technologists and

13:19

founders of our lifetime, Demis.

13:22

He called you while you were still at Oxford. And

13:25

then your company, Vision Factory

13:27

and DeepMind were both acquired by

13:29

Google back in 2014, around

13:31

the same time. And that's when

13:33

the two of you started

13:35

to work together now for over

13:37

10 years. What was it like

13:39

or what has it been like to work with Demis?

13:42

Yeah, I mean, Demis is

13:44

an incredible person, you

13:47

know, a real character and

13:49

a real visionary. And,

13:52

you know, also amazingly human

13:54

and relatable. And I think

13:56

that that really inspires people.

13:58

So, you know, it only

14:00

takes a five minute conversation

14:02

to, for him to sort

14:04

of really bleed out the

14:06

depth of ambition that he

14:08

thinks about. And

14:11

just the immediacy of

14:13

the potential to step

14:16

towards these ambitions. So

14:18

I think he has

14:20

this great ability to

14:22

inject a lot of

14:24

energy into a group

14:27

of very smart people,

14:29

get people to see beyond what's

14:32

right in front of them. I

14:35

remember moments sitting or

14:37

standing in the lobby

14:39

of one of the early Deep Mind offices. I

14:41

think this was the, it was a

14:43

toast. We were a celebration

14:45

we were having for the first nature paper from

14:47

Deep Mind. And

14:49

Demis was saying, you know, this

14:51

is actually just going to

14:53

be the first of dozens of

14:55

nature papers. And at the

14:58

time this was the first, basically the

15:00

first machine learning paper in nature. This was

15:02

the Atari DQN paper. And

15:04

the prospect of dozens of nature

15:06

papers, you know, it seems bit

15:08

far fetched and actually he went further

15:10

and said, and we're going to be

15:12

winning no prizes as a result of

15:15

this. And that was 10 years

15:17

ago. Yeah, that's incredible. the

15:19

forethought that he has. He's got what

15:21

I call like one of these roll out

15:23

minds. Maybe it comes from all of his experience

15:25

playing chess, but it's he's always, you know,

15:27

rolling out into the future. What are

15:29

the steps now that are going to lead, you know,

15:31

to this big ambition? So

15:34

yeah, it's been it's been fantastic. I've been

15:36

working with him for about 10 years now. still

15:40

work really closely together on isomorphic

15:42

labs. And the

15:44

ambition is as big as ever. It's so

15:46

interesting to hear that you had this ambition

15:48

and that he had this ambition from

15:50

the very start. And it's

15:52

incredible that it's played out that way. Well,

15:55

I'd love to talk a little bit

15:58

about isomorphic. You're now embarking on one

16:00

of the most ambitious missions of our

16:02

generation to reimagine drug discovery and drug

16:04

development with AI. Everything

16:06

goes right and you realize your

16:08

vision for isomorphic. What does the world

16:11

look like? Yeah,

16:13

you know, we think really

16:15

big isomorphic. We want

16:17

to be solving all

16:19

diseases here and genuinely that

16:21

scale. And the

16:23

point is that this technology

16:26

that we're building and AI

16:28

as a whole field is

16:30

going to be completely transformative. in

16:33

how we understand biology, in

16:36

our ability to manipulate

16:38

and craft chemistry to modulate

16:40

that biology. So

16:42

we really think about a

16:44

future where we are solving all

16:47

diseases where AI is not

16:49

just helping us discover and create

16:51

and design new therapeutics, but

16:53

also just understand so much more

16:55

about our biological world, about

16:57

how our you know cells

16:59

are working and what are

17:02

the root causes of disease and

17:04

therefore opening up new pathways

17:06

that we can think about modulating.

17:09

So we have set up

17:11

the company from day

17:13

one to really go after

17:15

this big ambition. This

17:17

isn't about developing

17:19

therapeutics for a particular

17:21

indication or a

17:23

particular target. It's really

17:25

thinking about how do we create

17:27

a very general drug design engine

17:30

with AI, something that we can

17:32

apply to not just a single

17:34

target or even a single modality,

17:37

but we can apply this again and

17:39

again across any different disease area. And

17:41

that's what we're stepping towards in the moment. How

17:44

does setting out with

17:46

this ambition of being

17:48

general change how you

17:50

built in practice from

17:52

day one? Yes, a

17:55

good question. When

17:57

I think about some of

17:59

the status quo of AI and

18:01

drug design, there's been a

18:03

lot of use of machine learning

18:05

models in chemistry and biology,

18:07

but I would call them a

18:09

lot of the first generation

18:11

of this sort of application to

18:13

be more local models where

18:15

you might have some data about

18:17

a particular target or about

18:20

how a particular class of molecules

18:22

is behaving and you'll fit

18:24

a small multi -layer

18:26

MLP against this data to

18:28

help you generate some predictions

18:30

that lead to your next

18:32

round of design. This

18:36

is the complete opposite approach

18:38

of what we were trying to

18:40

do. So from day one,

18:42

we were setting out to create

18:44

models that generalize across chemistry

18:46

and across target space. So,

18:48

you know, and a key example

18:50

of this is something like Alpha Fold

18:52

and Alpha Fold 3 where This

18:55

is a model that you can apply

18:57

to a whole different host of targets.

18:59

You can apply it to any protein

19:01

in the proteome, in the universe of

19:03

proteins. You can apply

19:05

it to any small molecule that you

19:07

can think of designing without needing

19:09

to fine -tune it, without needing to

19:11

fit any local data. And so you

19:13

can imagine that it completely changes

19:16

the way that chemists can use these

19:18

models if you don't need to

19:20

be adapting this model to every single

19:22

application. every

19:24

single one of our internal research projects.

19:26

And by the way, when I think

19:28

about what we're going to need to

19:30

get this breakthrough drug design engine that

19:32

we've been building, we

19:35

need like half a dozen alpha

19:37

folds. Alpha fold is just part

19:39

of the story. So

19:41

from day one, we've been

19:43

setting up these internal research programs

19:45

going after these half dozen

19:47

problems. We've had significant

19:50

breakthroughs, obviously in alpha folds and structure

19:52

prediction, but also in other key

19:54

areas. And

19:56

in all of these,

19:58

these models are general. They

20:00

can be applied to any target. And

20:03

then what we're finding actually, they can be

20:05

applied to any modality or lots of

20:07

different modalities. Yeah. So that's the first time

20:09

I've heard you say half a dozen

20:11

alpha folds. Can you share a little bit

20:13

more about what that means? Yeah.

20:15

So. Alpha -fold was obviously a

20:17

massive breakthrough in understanding biomolecular

20:19

structure. So what is the structure

20:21

of proteins? And now with

20:24

Alpha -fold, three structure proteins with

20:26

small molecules and things like DNA

20:28

and RNA. That's

20:30

a fundamental step change. It allows

20:32

us to get experimental level accuracy

20:35

of a really core concept of

20:37

biochemistry that unlocks a whole bunch

20:39

of thinking and design work for

20:41

chemists. My

20:44

comment here is actually we're probably

20:46

going to need something like half a

20:48

dozen more of these sort of

20:50

breakthroughs. They're sort of getting to experimental

20:52

level accuracy of different core concepts

20:54

of biology and chemistry. To

20:56

be able to put this

20:58

together into something that's really

21:00

transformative for drug design. Drug

21:03

design is really, really hard. It's not

21:05

just a single problem. It's not just

21:07

about understanding the structure of a protein.

21:10

It's not even just about designing a

21:12

molecule that will modulate that protein in

21:14

the way that you want. You want

21:16

this molecule to be able to ideally

21:18

be taken as a pill and go

21:20

through the body and be absorbed in

21:23

the right way and reach the right

21:25

cell type and actually go into the

21:27

cell and not be broken down by

21:29

the liver in a certain way. So

21:31

there's just so much complexity to hold

21:33

on to as a drug designer. And

21:36

each one of those is like

21:38

an alpha fold level style breakthrough that

21:40

we've been creating. So interesting. Well,

21:43

I've also heard you use the words

21:45

a holy grail model for drug design

21:47

and agents for science. Can you

21:49

explain a little bit more about what you mean? Yes.

21:51

So some of these research areas

21:53

that we've been going after, predicting

21:56

structure and properties of

21:58

these molecules and how all

22:01

of these biomolecules interact

22:03

and play out over time.

22:05

These really are sort of

22:07

holy grail. predictive problems for

22:09

drug design. And

22:11

we've made some incredible breakthroughs there,

22:14

which have really stunned our

22:16

chemists and step changed how we

22:18

do drug design internally, Iso. But

22:21

what I think a really interesting

22:24

thing to think about is that

22:26

you could create the best possible

22:28

predictive model of the world, like

22:30

an experimental level, even

22:32

better than experimental level model. to

22:35

predict a particular property about a molecule,

22:37

for example, to be able to

22:39

predict the outcome of a real experiment. See,

22:41

we can have a whole suite of those,

22:43

but that still wouldn't solve drug design. And

22:48

the way to think about this is, you

22:51

know, there's this number 10 to

22:53

the power of 60, which is

22:55

perhaps all of the possible drug

22:58

-like molecules that you could, that

23:00

could exist. you

23:02

know, that's maybe, that's maybe, that's maybe,

23:04

you know, a bit, you know, takes

23:06

into account a lot of things. So

23:08

we could even reduce that by 20

23:10

orders of magnitude, get to 10 to

23:12

the 40. That's still a lot

23:14

of things. And even if you

23:16

had the best predictive models in the

23:18

world, so let's say you could screen

23:20

a billion different molecules, you could go

23:22

and test a billion different molecules, that's

23:24

10 to the nine. So, you know,

23:26

now we're still like 10 to the

23:28

31 molecules left on the table. So

23:31

even with the best predictive models,

23:33

you're still not even scratching the surface

23:35

of molecular space that you should

23:37

be exploring. And this is

23:39

why we need to go

23:41

beyond just predictive models of experiment,

23:43

but also models like generative

23:45

models, like agents that can

23:47

actually navigate that whole 10 to

23:49

the 40, 10 to the 60

23:51

space. That's so interesting. Using our

23:53

predictive models, obviously, to understand how

23:55

to navigate that. But so we

23:57

don't have to exhaustively search because

23:59

we can never exhaustively search the

24:01

whole universe of molecules. If

24:03

that makes sense, just in the

24:05

same way that AlphaGo couldn't exhaustively search

24:07

all of the possible go moves, unlike

24:10

chess, where you could exhaustively search all

24:13

possible chess moves. But yeah, molecule designs

24:15

much more like go than it is

24:17

like chess. So

24:19

that's where generative models come

24:21

into play. Agents. that

24:23

utilize generative models, that utilize

24:25

search techniques as well as

24:28

these amazing predictive capabilities to

24:30

really open up the entirety

24:32

of molecular space. Now,

24:34

to me, it's actually still amazing

24:36

that even without AI, we managed to

24:38

find drugs in this 10 to

24:40

60 space, 10 to the 40 space.

24:44

It just says that actually there's probably a

24:46

lot of redundancy. There's a lot of

24:48

potential designs. If you think about

24:50

a particular disease indication, a particular target,

24:54

There should be many designs that exist

24:56

that would be good for that and

24:58

would be the right sort of product

25:00

profile for this therapeutic. And

25:03

I think the real potential

25:05

here is for these generative models,

25:08

these agents as well, to

25:10

be able to search through this

25:12

space and really uncover that

25:14

whole potential design space. That's so

25:16

interesting. I think in very

25:18

simplistic Lehmann terms, you're both... modeling,

25:21

learning and modeling the game

25:23

and trying to build the best

25:25

player to solve different types

25:27

of games. Yeah. So it's, I

25:29

mean, you know, I'm incredibly

25:32

biased by games. been playing

25:34

video games since I was a kid, grew

25:36

up in that world. But,

25:38

you know, that's exactly how I

25:40

think about it. We've got to

25:42

be creating our world models, our

25:44

models of the biochemical world, our

25:47

biological world. And

25:49

then we don't stop there. We

25:51

actually then need to be creating

25:53

agents and generative models that can

25:55

work out how to explore, how

25:57

to traverse that, and to basically

25:59

uncover these amazing needles in the

26:01

haystack in chemical space, which could

26:03

be life -changing therapeutics for so many

26:06

millions of people. I love that.

26:08

That is our punchline today. So

26:11

Alphapole 3 is truly groundbreaking. You've

26:14

taken us from being able to model

26:16

just the structure of a protein to now

26:18

being able to model the structure of

26:20

all molecules and their interactions with each other.

26:23

Can you share a little bit about how we

26:25

should think about that in terms of the

26:28

impact in accuracy, in speed and

26:30

efficiency, and also potentially in

26:32

being able to explore problem

26:34

spaces that we couldn't solve

26:36

before this? Yeah, so, yeah,

26:39

AlphaFol2 was, you

26:42

know, the biggest breakthrough,

26:44

right? To be able to understand

26:46

the structure of proteins, and

26:48

then there was something called alpha fold

26:50

to multimer, which then allows you to understand

26:52

not just the structure of proteins by

26:55

themselves, each individual protein, but the structure of

26:57

proteins as they come together and what

26:59

we call complexes, so how these proteins fit

27:01

together. That

27:03

opens up and helps us answer

27:05

a lot of questions in biology,

27:07

but there's still a big hop

27:09

to designing therapeutics. And one

27:11

of the big classes of therapeutics

27:13

is what's called small molecules. So

27:16

these are molecules that are not

27:18

proteins. These would be things like caffeine

27:20

or paracetamol, things that more often

27:22

you can take as a pill. And

27:26

the way that these therapeutics work with

27:28

these small molecules is that they go

27:30

through the body, they go into the

27:32

cell, and they actually come and attach

27:34

themselves to these proteins. These

27:37

proteins... the fundamental building blocks of

27:39

life. They form these molecular machines

27:41

by interacting with other proteins And

27:43

so you can you can imagine

27:45

that if you have another molecule

27:47

your drug that comes in and

27:49

attaches itself to a protein over

27:51

here Then it might disrupt the

27:53

ability for that protein to interact

27:55

with another protein one of its

27:57

normal machine and day -to -day life

27:59

And so you're modulating the function

28:01

of that protein with this small

28:04

molecule and that's the essence of

28:06

drug design and how

28:08

therapeutics work. And so you

28:10

can imagine as a chemist, your

28:12

job, a drug designer, you're trying

28:14

to design a small molecule that's going

28:16

to fit to this protein over

28:18

here and disrupt how it normally functions,

28:21

or in some cases, enhance how

28:23

it normally functions. And so

28:25

it'd be really helpful to understand

28:27

how this small molecule interacts with

28:29

the protein, what's the structure that

28:31

it might make, what are the

28:33

interactions, these literally physical interactions that

28:35

are being made. And

28:38

so that really inspired the

28:40

creation of AlphaFol3 where now

28:43

we have a model that

28:45

not only predicts the structure

28:47

of proteins, but how these

28:49

proteins interact with small molecules,

28:51

also other fundamental molecular machine

28:53

building blocks, things like DNA

28:55

and RNA. And

28:58

this basically opens up the ability to

29:00

structurally understand, which is a core part

29:02

of drug design, small

29:05

molecules. It opens

29:07

up new classes of targets. You

29:10

know, there are things like transcription

29:12

factors, which are proteins that sit on

29:14

DNA and read DNA. And you

29:16

can imagine now trying to design a

29:18

small molecule to change or disrupt

29:20

the function of something like that. And

29:22

so to do that, you'd really

29:24

want to be able to see literally

29:26

in 3D how this all looks.

29:28

And if I make changes to my

29:30

little molecule, how will that change

29:32

the way it interacts with this protein

29:34

and this biomolecular system? So Alpha

29:36

Fold 3 is now very, very accurate.

29:39

It allows us to answer a lot of

29:41

these questions purely in silico or purely

29:43

on the computer where before you would have

29:45

to go to the lab, literally crystallize

29:48

this stuff. This can take six months. It

29:50

can take years. Sometimes it's even impossible. Now

29:53

at ISO, our drug designers

29:56

are literally sitting with their laptop.

29:58

browser -based interface, being able to

30:00

understand, make changes to their

30:02

designs, and see the impact of

30:04

that. Incredible. So

30:06

there are a couple

30:08

of interactions that AlphaVol3 is

30:11

focused on, proteins in

30:13

nucleic acids, proteins in ligands,

30:15

and antibody to antigen.

30:17

Can you give us some

30:19

good examples of the

30:21

impact that AlphaVol3 now has

30:23

on the interaction of

30:25

these different types of proteins

30:27

and molecules? Yeah, so

30:29

protein and ligands, that's the same as

30:31

protein and small molecules. So those two

30:33

terms, ligands and small molecules are synonymous.

30:36

That allows us to understand how

30:38

small molecule drugs interact. Then

30:41

we can think

30:43

about protein -protein

30:45

interactions. There's

30:47

a whole class of therapeutics called

30:49

biologics. These are things like antibodies.

30:53

That allows us to understand how they might

30:55

interact with our targets. opens

30:57

up new modalities. And

31:00

that also encapsulates the

31:02

sort of the antibody anti

31:04

-gen interface. So if you're

31:06

designing an antibody, you

31:08

want to understand how your antibody

31:10

design is going to interact with

31:12

the protein surface there. So it's

31:14

the same model that we can

31:16

use across all of these different

31:18

applications. What are the

31:20

nuances of training a model like AlphaFol

31:22

-3? And what are the benefits of

31:24

using a diffusion -based architecture? Yes,

31:27

a great question. There were

31:29

a lot of challenges we had to overcome to get

31:31

AlphaFold 3 to work. One

31:33

of the most interesting things

31:35

was actually just how do we

31:37

take something like AlphaFold, which

31:39

was only working with proteins, and

31:41

then input these new modalities,

31:43

these new data types of RNA,

31:45

DNA, small molecules. So

31:47

we had to work out how to tokenize, not

31:49

just proteins, which we kind of knew how to

31:52

do, but how to tokenize then DNA, how

31:54

to tokenize small molecules. For things

31:56

like DNA and RNA, that's a little bit more

31:58

obvious. We could tokenize in the

32:00

bases. But then for

32:02

small molecules, we would really go to, we

32:05

tried a whole bunch of different

32:07

stuff. It really ended up that this

32:09

atomic resolution tokenization worked super well. And

32:13

then you have the question of, okay, how do

32:15

you actually predict

32:17

the structure of this

32:19

mixture of different molecule

32:21

types. You couldn't use

32:23

the same framework as

32:25

AlphaFol2 and this is

32:28

where diffusion modeling just

32:30

really shunned. Here

32:33

we could actually model

32:35

every single individual atom and

32:37

the coordinate of every

32:39

atom individually and have a

32:41

diffusion model be producing

32:43

those 3D coordinates and the

32:45

tokenization that we talked

32:47

about is conditioning the inference

32:49

of that diffusion process.

32:51

So interesting. And this was

32:53

a huge breakthrough. So,

32:55

you know, we're talking about

32:57

on our leaderboard just

32:59

a massive step change, particularly

33:01

in small molecule protein

33:03

interaction accuracy. It was a

33:05

massive step change and

33:07

something that really unblocks the

33:09

rest of the project.

33:11

Wow. So data

33:13

compute and algorithms. We

33:15

know those three are important in

33:17

all other adjacent fields. But

33:19

I was surprised to read an interview

33:22

with Demis where he shared that we're

33:24

not data constrained in biology. Can

33:26

you share your point of view on that? You

33:28

know, I think it doesn't matter what field

33:30

of machine learning you're in, you're going to feel

33:32

some data constrained. And

33:34

I think the point here from

33:37

Demis is that it's not a

33:39

real bottleneck, as in we can

33:41

make progress. with the data that

33:43

is out there, that the data

33:45

we can generate and real progress

33:47

can be made. It's

33:50

not, we've got to

33:52

sit and wait 50 years for the world

33:54

to generate data before we can actually make

33:56

impact here. No, we're not seeing that at

33:58

all. Modeling spaces

34:00

where the data has been

34:02

sitting around for years,

34:04

that we can see that

34:06

we can make really

34:08

substantial progress beyond anything that

34:10

people have experienced before. Now,

34:13

does that mean there's no opportunity for

34:15

data in biology? Absolutely not. It's

34:17

going to be a fundamental

34:19

part of how we continue

34:21

to develop these models and

34:23

these systems will be what

34:25

data we go out and

34:28

generate. And

34:30

there, I think, there's just a massive

34:32

opportunity. In my mind, the

34:36

data The data

34:38

for machine learning in biology hasn't actually

34:40

been created yet. Yes. Yes, there's a lot

34:42

of historical data, but there's a huge,

34:44

but that historical data hasn't been created for

34:47

the purposes of machine learning. And

34:49

so when you're going out and thinking, how

34:51

do I create data to actually train my model?

34:53

You're thinking in a very different way to

34:55

how people have gone out and generated data in

34:57

the past. And that there's a big opportunity

34:59

there to explore. What kind of

35:01

data do you think we're missing here right

35:03

now? And do we think, do you

35:05

think that we need? Anything

35:07

in synthetic data? Yes,

35:09

so I'm a massive fan of

35:11

synthetic data actually I have been

35:13

for since the very beginning of

35:16

my career where You know we

35:18

would I was generating synthetic text

35:20

data Just to overcome the fact

35:22

that you know I was a

35:24

PhD student with access to a

35:26

couple of thousand images and Google

35:28

had millions and millions of images

35:30

and so instead I just generated

35:32

tons and tons of synthetic data

35:34

and that unblocked things and We're

35:36

seeing the same thing in the,

35:39

especially the chemistry space where we

35:41

have good theory. We actually

35:43

know a lot about physics. We

35:45

know we have the theory

35:47

of quantum chemistry and quantum mechanics

35:49

and we can create simulators

35:51

out of that. We can

35:53

approximate that and create more scalable

35:56

molecular dynamic simulations. This

35:58

gives the basis for a whole

36:00

host of synthetic data. Then

36:02

we have the models themselves that Especially

36:05

we have generative models. This

36:07

can actually generate data that

36:09

we can use scoring systems

36:11

to help really enhance the

36:13

information content of this data.

36:16

But I think one of the big

36:18

open spaces will be on what's

36:20

called in vivo data. So

36:23

data that you would normally measure

36:25

on a real animal, something like a

36:27

mouse or a rat. There's

36:30

some historical data on that, but

36:32

you can't generate. tons of

36:34

that you can't really generate any

36:36

at all, right? So then there's a

36:38

big opportunity to look to new

36:40

data generating technologies. There are some incredible

36:42

people doing things like organoids on

36:44

a chip. So

36:47

ways of starting to measure

36:49

things that you would normally measure

36:51

on a real animal, but,

36:53

you know, completely on a chip.

36:55

So, you know, I think

36:57

there's gonna be a whole host

36:59

of like new breakthroughs in

37:02

data generating. technology in biology

37:04

and chemistry that's going to have big

37:06

impact on how we think about modeling

37:08

that world as well. Are you working

37:10

on any of that internally or

37:12

are you hoping that other players can

37:14

fill in some of that gap? So

37:17

internally, we actually don't have

37:19

any of our own labs

37:21

in isomorphic labs, but we

37:23

work with a whole bunch

37:25

of other companies. We

37:28

generate a lot of data ourselves, a

37:31

lot of proprietary data. We've seen an amazing impact

37:33

of that. It makes a lot of sense. So

37:36

there's a point of view that

37:38

modeling structure of molecules and modeling

37:41

their function and the modulation function

37:43

is very important, but not necessarily

37:45

always the limiting factor in drug

37:47

development. What's your point of view

37:49

on that? Yeah, as

37:51

I touched on one before, drug design

37:53

is really, really complex. And that's

37:55

before you even get to drug development,

37:57

which is where you take those

37:59

designs and you start putting them into

38:02

real people, clinical trials. There

38:04

are so many bottlenecks throughout

38:06

this whole design and development space.

38:10

Drug development is how

38:12

do we start

38:15

to approach clinical trials?

38:17

How should we test these drugs out in people? How

38:19

can we do this in a really timely manner, but

38:22

still a really safe manner? There's

38:25

a lot of bottlenecks there that I

38:27

think the industry as a whole. We

38:29

will need to work out

38:31

how to innovate in that space,

38:33

especially as our predictive models

38:35

of how these molecules will interact

38:37

with people, how toxic they

38:40

will be. As these predictive

38:42

models get better and better, we will

38:44

have to change the way that

38:46

we approach clinical trials to really make

38:48

use of that. Ultimately, to get

38:50

therapeutics into the hands of patients who

38:52

really desperately need them. Even

38:54

in the design of

38:57

molecules themselves as we talked

38:59

about before it's not just understanding

39:01

the structure of these molecules

39:03

it's not even just understanding how

39:05

these molecules change the function

39:07

of these proteins but we need

39:09

to understand how these molecules

39:12

change the function of pretty much

39:14

every single protein in our

39:16

body right because if we take

39:18

this as a pill it's

39:20

going to go everywhere and that's

39:22

the major cause of toxicity

39:24

is when Yes, you've designed

39:26

this amazing molecule that like perfectly

39:28

modulates your specific target that you know

39:31

is key to your disease But

39:33

also affects other things but it also

39:35

affects other things now, of course

39:37

you do a lot of screening to

39:39

protect against that but The more

39:41

we can predict that the better What's

39:44

really exciting from my perspective is

39:46

if we're creating these general models that

39:48

understand how this molecule interacts with

39:50

this target But also any other target

39:53

then why can't we just use that same

39:55

model to understand how these molecules interact

39:57

with the rest of our body? Right, so

39:59

interesting. So what is

40:01

now possible with AlphaFold 3 for

40:03

drug designers? How are you

40:05

using it internally? So AlphaFold

40:07

3 gives our drug designers

40:10

the ability to understand how

40:12

their molecule designs really interact

40:14

with this protein target. And

40:16

this is the target of

40:18

disease. And so our

40:21

drug designers can make changes to

40:23

the design and then see instantly

40:25

how that changes the way that

40:27

this molecule physically interacts with the

40:29

protein target. That's really,

40:31

really powerful. Before

40:33

our third, you would be completely

40:36

blind to this. You wouldn't actually

40:38

probably know how your molecule is

40:40

interacting with your protein. You'd

40:42

be using your best intuition, maybe

40:44

somewhere down the line in the

40:46

drug design project, you would get

40:48

your structure crystallized with a particular

40:50

design. That means going out to

40:52

a real lab six months later,

40:55

if you're lucky, getting a

40:57

resolved 3D structure. But

41:00

even then, that's just the 3D structure of

41:02

a single design, not every single change

41:04

that you make. Yeah. So

41:06

Alpha 3 completely changes the way

41:08

chemists can do this design work. But

41:10

I was stressed. nowhere

41:13

near as far as we want to go. Because

41:15

it's not just about what these molecules

41:17

look like in terms of interacting. We

41:19

actually want to know how strongly these

41:21

molecules interact with this protein. We

41:24

want to know other properties of

41:26

these molecules. We want to understand

41:28

how the way that these molecules

41:30

interact with this protein and how

41:32

that changes the fold or the

41:35

confirmation of the protein, how that

41:37

changes the function of the protein,

41:39

how it might actually change the

41:41

dynamics of the cell. There were

41:43

so many questions and these are

41:45

these other alpha fold like breakthroughs

41:48

that we're working on that also

41:50

go, you know We have created

41:52

incredible models for that our chemists

41:54

are using in this design process

41:56

interesting So you're designing some drugs

41:58

internally what targets and programs are

42:01

you focused on? So we have

42:03

a really exciting internal program of

42:05

drug design projects. These are focused

42:07

on immunology and oncology We've

42:09

been making some incredible progress there.

42:11

It's been really exciting to see

42:13

especially how these models have transformed

42:15

the way that we're actually approaching

42:18

drug design on these these programs.

42:20

You're also working with Eli Lilly

42:22

and Novartis and recently you announced

42:24

an expansion with Novartis's partnership. Can

42:26

you share a little bit about

42:28

what these partnerships look like? Yes,

42:31

so we we signed these

42:33

initial partnerships two

42:35

partnerships, one with Eli Lilly, one

42:37

with Novartis. That was fantastic. They

42:39

brought some really, really challenging problems

42:41

to us. I think it's no

42:43

secret that, you know, the sort

42:46

of targets that, for example, Novartis

42:48

brought to us, these are these

42:50

are sort of targets that, you

42:52

know, the field and Novartis, for

42:54

example, been working on for, you

42:56

know, 10 years plus. So

42:59

these aren't sort of

43:01

old. We'll try things out,

43:03

problems. These are for real

43:05

hard things. Last

43:08

year was an amazing year, both

43:10

for our internal projects, but also

43:12

for these partner projects to really

43:14

see how well these models are

43:16

working. It's allowed

43:18

us to really uncover new chemical

43:21

matter, working on new ways

43:23

to modulate these targets that people

43:25

have worked on for a long time.

43:28

It's been amazing to see this

43:31

new deal which has expanded on the

43:33

Vartis collaboration, which I think is a

43:35

real testament to some of the success

43:37

of the early days of these partnerships.

43:40

Congratulations. I think it's an incredible

43:42

milestone, especially just one year in. Yeah.

43:44

So I'd love to talk a little bit

43:46

about the team. You've built a truly

43:48

excellent team composed of the highest caliber talent

43:50

across many different fields, AI,

43:52

chemistry, biology, and

43:54

you've also brought outsiders into the

43:56

field to help question traditional thinking. Can

43:58

you share a little bit about how

44:00

you thought about this? Yeah,

44:03

so the space of

44:05

AI for drug design

44:07

hasn't really existed for

44:09

very long. So

44:11

the chances of finding a

44:13

world expert at drug design

44:15

who's also a world expert

44:18

and machine learning or deep

44:20

learning is basically zero. Just

44:23

because these fields, these fields haven't

44:25

coexisted for long enough. I genuinely

44:27

think about a new sort of

44:30

a field of science that ISO

44:32

is breeding because we are, you

44:34

know, we have these people who

44:36

really live and breathe the intersection

44:38

of this. So,

44:40

you know, because because we

44:42

can't hire these people, you know,

44:44

I really think about how do

44:46

we bring the world experts at

44:48

drug design and medicinal chemistry? and

44:51

the world experts at machine learning and

44:53

deep learning, and get these

44:55

incredible people sitting side

44:57

by side, because it's

44:59

not just enough to have

45:01

these amazing people sitting in their

45:03

isolated teams. We

45:06

need people sitting side by side,

45:08

speaking each other's languages with

45:10

a lot of empathy, a

45:12

lot of curiosity, curiosity

45:14

to understand this new science,

45:16

to really build intuitions in

45:18

your own language. And

45:20

we've seen just such amazing

45:23

things come out of this

45:25

dynamic where you really have

45:27

a generalist machine learner who

45:29

doesn't know anything about chemistry

45:31

or biology, start to come

45:33

in and understand the problems

45:36

of a medicinal chemist and

45:38

a drug designer. And

45:40

when I think about even

45:42

hiring machine learners and machine learning

45:44

scientists and engineers for the

45:46

research that we're doing, I'd

45:49

say, you know, 60, 70,

45:51

80 % of the people on

45:54

our team have no prior knowledge

45:56

of chemistry or biology, maybe,

45:58

you know, high school or university

46:00

level. And that can

46:02

actually be a real asset because

46:04

you come in sort of a

46:06

little bit naive. And

46:08

as long as you're curious, I think

46:10

one of the key things is

46:13

asking, you know, the curious questions, asking

46:15

this like stupid questions. And

46:17

then that allows us to

46:19

come at the problems from

46:21

first principles. It almost allows

46:23

us to break through the

46:26

dogma of previous experience and

46:28

how people traditionally approach these

46:30

problems. We can think ground

46:32

up from scratch. And that's

46:34

a lot of the mentality of how

46:36

we think about creating these research breakthroughs.

46:38

A little naive and highly curious and

46:41

high agency is a very good thing.

46:43

Yes, exactly, exactly. So in

46:45

November last year, you also made a very

46:47

big move in launching the AlphaFold server, which

46:49

releases code and model weights for

46:51

academic use. Can you share a little

46:53

bit about why? Yeah,

46:55

so AlphaFold has a long,

46:57

long lineage of being open

47:00

for this academic and scientific

47:02

use. And it

47:04

was really important with this

47:06

latest breakthrough of AlphaFold

47:08

3 that we make sure

47:10

that this scientific community has

47:13

access to this functionality because, you

47:15

know, yes, outfall three is going to

47:17

be incredibly useful for drug design.

47:19

It already is. But it's

47:21

also useful for, you know,

47:23

many other areas of fundamental biology

47:26

and just understanding biology. And

47:28

people are using these people are

47:30

using outfall three server and

47:32

modeling in very, very creative ways.

47:35

So, you know, it's very important for

47:37

us to make sure that there is

47:39

that sort of free use for non -commercial

47:41

academic work. And it's been

47:43

incredible to see the take up of that

47:46

and the use of the server. Let's

47:48

talk a little bit about the future. Can

47:50

you give us a tease of what else is

47:52

to come with AlphaFold? In

47:55

terms of structure prediction

47:57

as a problem, in

47:59

my mind, I want

48:01

to completely solve this. I

48:04

think AlphaFold 3 is a... fantastic step

48:06

on the way of that. There's a

48:08

significant breakthrough. But

48:11

it's not 100 % accuracy. What

48:14

does even 100 % accuracy mean in

48:16

this space? Like

48:18

with a lot of areas of

48:20

science, as you start to push the

48:22

boundaries, you see that the problem

48:24

opens up into even more problems. That's

48:27

the addictive part of doing science.

48:32

Alpha Fold 3 is a good example

48:34

of that, where as you start to

48:36

get these capabilities, you see

48:38

that actually there are even more

48:40

deeper problems that we want to be

48:43

working on and stepping towards. So

48:45

yes, understanding structure better and better and

48:47

more accurately is always going to

48:49

be interesting for us. But then it's

48:51

also not just necessarily about static

48:53

structure. So Alpha Fold 3 models these

48:56

crystal structures. which are

48:58

almost static crystallized versions

49:00

of these molecules, how these

49:02

molecules interact. But in

49:04

reality, we don't have crystals inside

49:06

of us. These molecules are in solution.

49:08

They're moving about the dynamic. So

49:10

you can think, OK, well, maybe

49:13

understanding the dynamics of these systems is

49:15

actually also going to be really

49:17

interesting. What does a

49:19

GPT -3 moment look like in AI

49:21

biology? And when do we get

49:23

there? So if I think about

49:25

GPT -3, This

49:27

is really a generative model. So

49:30

something that's generating text. And

49:32

the GPT -3 moment for

49:34

me was, you

49:37

know, crossing over that

49:39

boundary between, yeah,

49:41

we've got generative models of text and

49:43

they generate some stuff and it looks

49:45

like text, but I'm not convinced that

49:47

it's generated by a human. And

49:50

GPT -3 started to be that first

49:52

point where you're like... shit. This

49:54

kind of looks like a human. And

49:57

so this generative model is

49:59

actually recreating the distribution of

50:01

data that is trained on.

50:04

And what is a generative model? Generative

50:06

model is something that fits the manifold

50:08

of data that is trained on it.

50:10

So when I think about this applied

50:12

to biology, you can

50:14

think about these generative

50:17

models actually starting to

50:19

recreate that GPT -3

50:21

moment. recreate what things

50:23

would actually look like in reality. And

50:26

that's quite exciting because that means

50:29

that these models are spitting out things

50:31

that either they actually exist in

50:33

the world and we can kind of

50:35

validate that or maybe even discover

50:37

new things that exist in the world

50:39

or they could exist in the

50:41

world, which means that they could be

50:43

things that we could design or

50:46

manufacture or create. that would actually

50:48

be stable and work and exist in

50:50

our physical reality. And I

50:52

think the cool thing

50:54

about this in biology is

50:56

that, unlike with language,

50:58

where with language, when

51:00

it generates something human level quality,

51:02

we can understand that because it

51:04

is human derived. But

51:06

a lot of problems in chemistry and

51:08

biology, we even struggle to understand ourselves. And

51:10

so when we get to that GPT -3

51:12

moment, I think it will look a

51:14

lot less like GPT -3, but much more

51:16

feel a lot more like move 37 in

51:18

AlphaGo. Interesting. Where we're starting

51:21

to see things that are beyond

51:23

human understanding, but that do exist

51:25

in the real world, that exist

51:27

in our physical reality, but

51:29

are beyond sort of human comprehension.

51:31

Right. And that's just going to

51:33

be mind blowing. In fact, you

51:36

know, we're starting to see that

51:38

internally with our generative models, that

51:41

we're creating designs that a

51:43

human drug designer would say,

51:46

Hmm, I'm not so sure about that.

51:48

I much prefer this and then

51:50

you test it out in physical reality

51:52

and The generative model is correct

51:54

and the human is wrong. That's fascinating

51:57

I love the move 37 analogy

51:59

when the model starts to see elements

52:01

of creativity and exactly past the

52:03

human move 37 was this Amazing move

52:05

during the Alpha go games against

52:07

Lisa doll it was, you

52:09

know, the 37th move of the game

52:11

and it stunned the world, stunned the go

52:14

world because it was uninterpretable by human.

52:16

It looked like a mistake. No one had

52:18

ever played this move in the entirety of,

52:20

you know, thousands of years of human history

52:22

playing go. And it turned out

52:24

as you unrolled the game that this was

52:27

the critical move that allowed AlphaGo to beat

52:29

Lee Sedol in that match. And we're going

52:31

to see so much of that sort of

52:33

behavior coming out of these models, especially when

52:35

we're applying them to Things

52:37

outside of of native human

52:39

understanding like chemistry and biology.

52:41

Yeah, I love that also

52:43

our punchline today So when

52:45

we see our first AI

52:47

generated drug in clinic and

52:49

also in phase one two

52:51

and three trials so we're

52:53

making amazing progress on our

52:55

drug design programs and you

52:57

know the thing I think

52:59

about actually is as We

53:01

start to get a whole

53:03

bunch of these AI

53:06

designed assets, these molecules

53:09

get into clinical phase.

53:12

How can we actually start

53:14

to think about engaging in

53:16

that clinical development to get

53:18

these molecules to people as

53:20

fast and as safely as

53:23

possible because there's so much

53:25

unmedical need? So

53:28

yeah, here I think about

53:30

what are gonna be new

53:32

ways to engage with regulatory

53:34

bodies? What are going to be

53:36

new ways to incorporate our predictive models for

53:38

not only how this molecule works for the disease,

53:41

but how, as we talked about, how it

53:43

interacts with the rest of the body, the

53:45

types of toxicity it may induce.

53:48

I think there'll be a lot

53:50

of opportunities to think about just

53:53

streamlining and speeding up this process,

53:55

maybe even completely changing the way

53:57

we think about human clinical trials

53:59

as we, you know, our AI models

54:01

become so, you know, we can design

54:03

these molecules so much quicker in a much

54:05

more targeted manner with so much more

54:07

knowledge about how they work. Yeah. So

54:10

that'll change the game. But I think

54:12

we've got a long way to go as

54:14

an industry to really work out how

54:16

that changes. Yeah. Last question. As

54:18

isomorphic succeeds and potentially as a

54:20

whole field succeeds, what happens to

54:22

the traditional world of pharma? I

54:25

think they become, you know, in

54:27

some sense, pharma be

54:29

using AI. I think I think there's

54:31

no world when five years time you

54:33

will be designing a drug without

54:35

AI. That an inevitability.

54:39

It'll be like you know, trying

54:41

to do science without using

54:43

maths. AI will be this fundamental

54:45

tool for biology and chemistry.

54:47

It already is, at least in

54:50

isomorphics world, that

54:52

everyone will be

54:54

using. So it's not

54:56

going to be, Oh, is it pharma or it

54:58

AI? There's to be one and the same

55:00

in sense that the whole industry will adapt to

55:02

that. Yeah. Amazing. Max,

55:04

thank you so much for joining us today. This

55:06

was a fascinating conversation. been a pleasure. Thank you.