00:00:00.320 --> 00:00:01.607
[MUSIC]    

00:00:01.607 --> 00:00:02.572
[BOOK PASSAGE]

00:00:02.572 --> 00:00:08.160
PETER LEE: “Can GPT-4 indeed accelerate the 
progression of medicine … ? It seems like a tall  

00:00:08.160 --> 00:00:14.000
order, but if I had been told six months ago that 
it could rapidly summarize any published paper,  

00:00:14.000 --> 00:00:17.520
that alone would have satisfied me 
as a strong contribution to research  

00:00:17.520 --> 00:00:22.720
productivity. … But now that I've seen what 
GPT-4 can do with the healthcare process,  

00:00:22.720 --> 00:00:27.200
I expect a lot more in the realm of research.”

00:00:27.200 --> 00:00:28.288
[END OF BOOK PASSAGE]    

00:00:28.288 --> 00:00:29.120
[THEME MUSIC]    

00:00:29.120 --> 00:00:42.400
This is The AI Revolution in Medicine, 
Revisited. I’m your host, Peter Lee.    

00:00:42.400 --> 00:00:48.240
Shortly after OpenAI's GPT-4 was publicly 
released, Carey Goldberg, Dr. Zak Kohane,  

00:00:48.240 --> 00:00:53.680
and I published The AI Revolution in Medicine 
to help educate the world of healthcare and  

00:00:53.680 --> 00:00:58.240
medical research about the transformative 
impact this new generative AI technology  

00:00:58.240 --> 00:01:03.280
could have. But because we wrote the 
book when GPT-4 was still a secret,  

00:01:03.280 --> 00:01:10.720
we had to speculate. Now, two years later, what 
did we get right, and what did we get wrong?     

00:01:10.720 --> 00:01:15.360
In this series, we’ll talk to clinicians, 
patients, hospital administrators,  

00:01:15.360 --> 00:01:23.977
and others to understand the reality of AI 
in the field and where we go from here.      

00:01:23.977 --> 00:01:24.000
[THEME MUSIC FADES] 

00:01:24.000 --> 00:01:26.960
The book passage I read at 
the top was from “Chapter 8:  

00:01:26.960 --> 00:01:30.240
Smarter Science,” which was written by Zak.

00:01:30.240 --> 00:01:36.480
In writing the book, we were optimistic about 
AI’s potential to accelerate biomedical research  

00:01:36.480 --> 00:01:43.200
and help get new and much-needed treatments and 
drugs to patients sooner. One area we explored  

00:01:43.200 --> 00:01:48.640
was generative AI as a designer of clinical 
trials. We looked at generative AI’s adeptness  

00:01:48.640 --> 00:01:55.200
at summarizing helping speed up pre-trial triage 
and research. We even went so far as to predict  

00:01:55.200 --> 00:02:00.320
the arrival of a large language model that 
can serve as a central intellectual tool.

00:02:00.320 --> 00:02:04.560
For a look at how AI is impacting 
biomedical research today,  

00:02:04.560 --> 00:02:09.520
I’m excited to welcome Daphne Koller, 
Noubar Afeyan, and Eric Topol.

00:02:09.520 --> 00:02:15.200
Daphne Koller is the CEO and founder of 
Insitro, a machine learning-driven drug  

00:02:15.200 --> 00:02:18.720
discovery and development company 
that recently made news for its  

00:02:18.720 --> 00:02:24.000
identification of a novel drug target for 
ALS and its collaboration with Eli Lilly to  

00:02:24.000 --> 00:02:28.560
license Lilly's biochemical delivery 
systems. Prior to founding Insitro,  

00:02:28.560 --> 00:02:34.800
Daphne was the co-founder, co-CEO, and president 
of the online education platform Coursera.

00:02:34.800 --> 00:02:39.200
Noubar Afeyan is the founder 
and CEO of Flagship Pioneering,  

00:02:39.200 --> 00:02:43.360
which creates biotechnology companies 
focused on transforming human health and  

00:02:43.360 --> 00:02:48.160
environmental sustainability. He is also 
co-founder and chairman of the messenger  

00:02:48.160 --> 00:02:53.760
RNA company Moderna. An entrepreneur and 
biochemical engineer, Noubar has numerous  

00:02:53.760 --> 00:02:58.400
patents to his name and has co-founded 
many startups in science and technology.

00:02:58.400 --> 00:03:04.320
Dr. Eric Topol is the executive vice president 
of the biomedical research non-profit Scripps  

00:03:04.320 --> 00:03:10.000
Research, where he founded and now directs the 
Scripps Research Translational Institute. One  

00:03:10.000 --> 00:03:14.640
of the most cited researchers in medicine, 
Eric has focused on promoting human health  

00:03:14.640 --> 00:03:20.160
and individualized medicine through the 
use of genomic and digital data and AI.

00:03:20.160 --> 00:03:24.160
These three are likely to have an 
outsized influence on how drugs and  

00:03:24.160 --> 00:03:27.120
new medical technologies soon will be developed. 

00:03:27.120 --> 00:03:27.760
[TRANSITION MUSIC]

00:03:27.760 --> 00:03:32.680
Here’s my interview with Daphne Koller: 

00:03:32.680 --> 00:03:39.585
LEE: Daphne, I'm just 
thrilled to have you join us.

00:03:39.585 --> 00:03:42.200
DAPHNE KOLLER: Thank you for having 
me, Peter. It's a pleasure to be here.

00:03:42.200 --> 00:03:50.720
LEE: Well, you know, you're quite 
well-known across several fields. But  

00:03:50.720 --> 00:03:56.400
maybe for some audience members of this podcast, 
they might not have encountered you before.  

00:03:56.400 --> 00:04:02.080
So where I'd like to start is a question 
I've been asking all of our guests.

00:04:03.120 --> 00:04:07.200
How would you describe what you do? And 
the way I kind of put it is, you know,  

00:04:07.200 --> 00:04:12.160
how do you explain to someone like 
your parents what you do for a living?

00:04:12.160 --> 00:04:16.800
KOLLER: So that answer obviously 
has shifted over the years.

00:04:17.600 --> 00:04:22.720
What I would say now is that we are 
working to leverage the incredible  

00:04:22.720 --> 00:04:29.200
convergence of very powerful technologies, 
of which AI is one but not the only one,  

00:04:29.200 --> 00:04:34.240
to change the way in which we 
discover and develop new treatments  

00:04:34.240 --> 00:04:38.640
for diseases for which patients are 
currently suffering and even dying.

00:04:38.640 --> 00:04:43.360
LEE: You know, I think I've 
known you for a long time.

00:04:43.360 --> 00:04:47.746
KOLLER: Longer than I think 
either of us care to admit.

00:04:47.746 --> 00:04:52.320
LEE: [LAUGHS] In fact, I think I remember you 
even when you were still a graduate student.  

00:04:52.320 --> 00:04:57.680
But of course, I knew you best when you took up 
your professorship at Stanford. And I always,  

00:04:57.680 --> 00:05:03.360
in my mind, think of you as a computer scientist 
and a machine learning person. And in fact,  

00:05:03.360 --> 00:05:10.560
you really made a big name for yourself in 
computer science research in machine learning.

00:05:10.560 --> 00:05:14.640
But now you're, you know, leading one of the most  

00:05:14.640 --> 00:05:18.280
important biotech companies on 
the planet. How did that happen?

00:05:18.280 --> 00:05:23.120
KOLLER: So people often think that 
this is a recent transition. That is,  

00:05:23.120 --> 00:05:27.360
after I left Coursera, I looked around 
and said, “Hmm. What should I do next? Oh,  

00:05:27.360 --> 00:05:32.080
biotech seems like a good thing,” but 
that's actually not the way it transpired.

00:05:32.080 --> 00:05:38.720
This goes all the way back to my early 
days at Stanford, where, in fact,  

00:05:38.720 --> 00:05:44.560
I was, you know, as a young faculty member 
in machine learning, because I was the first  

00:05:44.560 --> 00:05:49.200
machine learning hire into Stanford's computer 
science department, I was looking for really  

00:05:49.200 --> 00:05:54.720
exciting places in which this technology 
could be deployed, and applications back then,  

00:05:54.720 --> 00:05:58.640
because of scarcity of data, 
were just not that inspiring.

00:05:58.640 --> 00:06:05.840
And so I looked around, and this was around 
the late ’90s, and realized that there  

00:06:05.840 --> 00:06:11.440
was interesting data emerging in biology and 
medicine. My first application actually was in,  

00:06:11.440 --> 00:06:16.640
interestingly, in epidemiology—patient tracking 
and tuberculosis. You know, you can think of it as  

00:06:16.640 --> 00:06:23.680
a tiny microcosm of the very sophisticated models 
that COVID then enabled in a much later stage.

00:06:23.680 --> 00:06:23.728
LEE: Right.

00:06:23.728 --> 00:06:29.120
KOLLER: And so initially, this was based almost 
entirely on just technical interest. It's kind of  

00:06:29.120 --> 00:06:34.160
like, oh, this is more interesting as a question 
to tackle than spam filtering. But then I became  

00:06:34.160 --> 00:06:39.440
interested in biology in its own right, biology 
and medicine, and ended up having a bifurcated  

00:06:39.440 --> 00:06:44.560
existence as a Stanford professor where half my 
lab continued to do core computer science research  

00:06:44.560 --> 00:06:49.520
published in, you know, NeurIPS and ICML. And 
the other half actually did biomedical research  

00:06:49.520 --> 00:06:55.040
that was published in, you know, Nature Cell [and] 
Science. So that was back in, you know, the early,  

00:06:55.040 --> 00:07:00.720
early 2000s, and for most of my Stanford 
career, I continued to have both interests.

00:07:00.720 --> 00:07:06.640
And then the Coursera experience kind of took me 
out of Stanford and put me in an industry setting  

00:07:06.640 --> 00:07:12.000
for the first time in my life actually. But then 
when my time at Coursera came to an end, you know,  

00:07:12.000 --> 00:07:16.880
I'd been there for five years. And if you look 
at the timeline, I left Stanford in early 2012,  

00:07:16.880 --> 00:07:20.800
right as the machine learning revolution 
was starting. So I missed the beginning.

00:07:20.800 --> 00:07:27.600
And it was only in like 2016 or so that, as 
I picked my head up over the trenches, like,  

00:07:27.600 --> 00:07:33.600
“Oh my goodness, this technology is going to 
change the world.” And I wanted to deploy that  

00:07:33.600 --> 00:07:40.560
big thing towards places where it would have 
beneficial impact on the world, like to make  

00:07:40.560 --> 00:07:42.000
the world a better place.  
 
LEE: Yeah. 

00:07:42.000 --> 00:07:47.360
KOLLER: And so I decided that one of 
the areas where I could make a unique,  

00:07:47.360 --> 00:07:53.600
differentiated impact was in really bringing 
AI and machine learning to the life sciences,  

00:07:53.600 --> 00:07:58.080
having spent, you know, the majority of my 
career at the boundary of those two disciplines.  

00:07:58.080 --> 00:08:02.520
And notice I say “boundary” with deliberation 
because there wasn't very much of an intersection.

00:08:02.520 --> 00:08:03.120
LEE: Right.

00:08:03.120 --> 00:08:06.160
KOLLER: I felt like I could 
do something that was unique.

00:08:06.160 --> 00:08:11.360
LEE: So just to stick on you for a little bit 
longer, you know, we have been sort of getting  

00:08:11.360 --> 00:08:17.440
into your origin story about what we call AI 
today—but machine learning, so deep learning.

00:08:17.440 --> 00:08:23.680
And, you know, there has always been a kind of 
an emotional response for people like you and  

00:08:23.680 --> 00:08:31.520
me and now the general public about their first 
encounters with what we now call generative AI.  

00:08:31.520 --> 00:08:38.680
I’d love to hear what your first encounter was 
with generative AI and how you reacted to this.

00:08:38.680 --> 00:08:47.680
KOLLER: I think my first encounter was 
actually an indirect one. Because, you know,  

00:08:47.680 --> 00:08:54.720
the earlier generations of generative AI 
didn’t directly touch our work at Insitro.

00:08:54.720 --> 00:09:00.480
And yet at the same time, I had always 
had an interest in computer vision.  

00:09:00.480 --> 00:09:05.920
That was a large part of my non-bio 
work when I was at Stanford.

00:09:05.920 --> 00:09:11.840
And so some of my earlier even presentations, 
when I was trying to convey to people back in  

00:09:11.840 --> 00:09:17.200
2016 how this technology was going to transform 
the world, I was talking about the incredible  

00:09:17.200 --> 00:09:21.120
progress in image recognition that 
had happened up until that point.

00:09:21.120 --> 00:09:26.960
So my first interaction was actually 
in the generative AI for images,  

00:09:26.960 --> 00:09:29.818
where you are able to go the other way …

00:09:29.818 --> 00:09:29.828
LEE: Yes.

00:09:29.828 --> 00:09:36.240
KOLLER: … where you can take a verbal 
description of an image and create—and  

00:09:36.240 --> 00:09:39.920
this was back in the days when the images 
weren't particularly photorealistic,  

00:09:39.920 --> 00:09:47.200
but still a natural language description to an 
image was magic given that only two or three  

00:09:47.200 --> 00:09:54.080
years before that, we were barely able to look 
at an image and write a short phrase saying,  

00:09:54.080 --> 00:10:03.720
“This is a dog on the beach.” And so that arc, 
that hockey curve, was just mind blowing to me.

00:10:03.720 --> 00:10:06.200
LEE: Did you have moments of skepticism?

00:10:06.200 --> 00:10:13.840
KOLLER: Yeah, I mean the early, you know, 
early versions of ChatGPT, where it was more  

00:10:13.840 --> 00:10:21.280
like parlor tricks and poking it a little bit 
revealed all of the easy ways that one could  

00:10:21.280 --> 00:10:25.360
break it and make it do really stupid things. 
I was like, yeah, OK, this is kind of cute,  

00:10:25.360 --> 00:10:30.960
but is it going to actually make a difference? 
Is it going to solve a problem that matters?

00:10:30.960 --> 00:10:35.440
And I mean, obviously, I think now 
everyone agrees that the answer is yes,  

00:10:35.440 --> 00:10:38.800
although there are still people who are like, 
yeah, but maybe it's around the edges. I'm not  

00:10:38.800 --> 00:10:43.280
among them, by the way, but ... yeah, 
so initially there were like, “Yeah,  

00:10:43.280 --> 00:10:47.520
this is cute and very impressive, but is it going 
to make a difference to a problem that matters?”

00:10:47.520 --> 00:10:52.160
LEE: Yeah. So now, maybe this is a good 
time to get into what you've been doing  

00:10:52.160 --> 00:10:56.080
with ALS [amyotrophic lateral sclerosis]. 
You know, there's a knee-jerk reaction from  

00:10:56.080 --> 00:11:03.360
the technology side to focus on designing 
small molecules, on predicting, you know,  

00:11:03.360 --> 00:11:08.678
their properties, you know, maybe binding 
affinity or aspects of ADME [absorption,  

00:11:08.678 --> 00:11:11.520
distribution, metabolism, and excretion], you 
know, like absorption or dispersion or whatever.

00:11:11.520 --> 00:11:17.040
And all of that is very useful, but 
if I understand the work on ALS,  

00:11:17.040 --> 00:11:23.240
you went to a much harder place, which is 
to actually identify and select targets.

00:11:23.240 --> 00:11:24.800
KOLLER: That’s right.

00:11:24.800 --> 00:11:30.880
LEE: So first off, just for the benefit 
of the standard listeners of this podcast,  

00:11:30.880 --> 00:11:34.440
explain what that problem is in general.

00:11:34.440 --> 00:11:39.520
KOLLER: No, for sure. And I 
think maybe I'll start by just  

00:11:39.520 --> 00:11:43.252
very quickly talking about the drug 
discovery and development arc, …

00:11:43.252 --> 00:11:43.268
LEE: Yeah. 

00:11:43.268 --> 00:11:48.000
KOLLER: … which, by and large, consists of 
three main phases. That's the standard taxonomy.

00:11:48.000 --> 00:11:53.920
The first is what's called sometimes target 
discovery or identifying a therapeutic hypothesis,  

00:11:53.920 --> 00:11:58.720
which looks like: if I modulate this target in 
this disease, something beneficial will happen.

00:11:58.720 --> 00:12:02.320
Then, you have to take that target and turn it 
into a molecule that you can actually put into  

00:12:02.320 --> 00:12:06.400
a person. It could be a small molecule. It 
could be a large molecule like an antibody,  

00:12:06.400 --> 00:12:11.680
whatever. And then you have that construct, that 
molecule. And the last piece is you put it into  

00:12:11.680 --> 00:12:18.400
a person in the context of a clinical trial, and 
you measure what has happened. And there's been AI  

00:12:18.400 --> 00:12:22.480
deployed towards each of those 
three stages in different ways.

00:12:23.040 --> 00:12:28.560
The last one is mostly like an efficiency gain. 
You know, the trial is kind of already defined,  

00:12:28.560 --> 00:12:31.920
and you want to deploy technology to 
make it more efficient and effective,  

00:12:31.920 --> 00:12:34.960
which is great because those 
are expensive operations.

00:12:34.960 --> 00:12:35.272
LEE: Yep.

00:12:35.272 --> 00:12:39.440
KOLLER: The middle one is where I would 
say the vast majority of efforts so far  

00:12:39.440 --> 00:12:45.440
has been deployed in AI because it is a nice, 
well-defined problem. It doesn't mean it's easy,  

00:12:45.440 --> 00:12:52.240
but it's one where you can define the problem. It 
is, I need to inhibit this protein by this amount,  

00:12:52.240 --> 00:12:59.120
and the molecule needs to be soluble and whatever 
and go past the blood-brain barrier. And you know  

00:12:59.120 --> 00:13:04.240
probably within a year and a half or 
so, or two, if you succeeded or not.

00:13:04.240 --> 00:13:08.880
The first stage is the one where I would 
say the least amount of energy has gone  

00:13:08.880 --> 00:13:13.120
because when you're uncovering a novel 
target in the context of an indication,  

00:13:13.120 --> 00:13:16.640
you don't know that you've been successful 
until you go all the way to the end,  

00:13:16.640 --> 00:13:21.120
which is the clinical trial, which is what 
makes this a long and risky journey. And  

00:13:21.120 --> 00:13:25.520
not a lot of people have the appetite 
or the capital to actually do that.

00:13:25.520 --> 00:13:31.360
However, in my opinion, and that of, I think, 
quite a number of others, it is where the  

00:13:31.360 --> 00:13:38.640
biggest impact can be made. And the reason 
is that while pharma has its deficiencies,  

00:13:39.200 --> 00:13:42.240
making good molecules is actually 
something they're pretty good at.

00:13:42.960 --> 00:13:46.720
It might take them longer than it should, 
maybe it's not as efficient as it could be,  

00:13:46.720 --> 00:13:51.200
but at the end of the day, if you tell them 
to drug A target, pharma is actually pretty  

00:13:51.200 --> 00:13:56.000
good at generating those molecules. However, 
when you put those molecules into the clinic,  

00:13:56.000 --> 00:14:00.480
90% of them fail. And the reason they fail 
is not by and large because the molecule  

00:14:00.480 --> 00:14:04.720
wasn't good. In the majority of cases, 
it's because the target you went after  

00:14:04.720 --> 00:14:08.960
didn't do anything useful in the context of 
the patient population in which you put it.

00:14:08.960 --> 00:14:13.360
And so in order to fix the 
inefficiency of this industry,  

00:14:13.360 --> 00:14:18.400
which is incredible inefficiency, you 
need to address the problem at the root,  

00:14:18.400 --> 00:14:23.120
and the root is picking the right targets to 
go after. And so that is what we elected to do.

00:14:23.120 --> 00:14:25.280
It doesn't mean we don't make 
molecules. I mean, of course,  

00:14:25.280 --> 00:14:28.480
you can't just end up with a target 
because a target is not actionable.  

00:14:28.480 --> 00:14:31.520
You need to turn it into a molecule. And 
we absolutely do that. And by the way,  

00:14:31.520 --> 00:14:35.182
the partnership with Lilly is actually 
one where they help us make a molecule.

00:14:35.182 --> 00:14:35.192
LEE: Yes.

00:14:35.192 --> 00:14:39.200
KOLLER: I mean, it's our target. It's our 
program. But Lilly is deploying its very  

00:14:39.920 --> 00:14:44.720
state-of-the-art molecule-making capabilities 
to help us turn that target into a drug.

00:14:44.720 --> 00:14:50.480
LEE: So let's get now into the 
machine learning of this. Again,  

00:14:50.480 --> 00:14:54.800
this just strikes me as such 
a difficult problem to solve.

00:14:54.800 --> 00:14:55.760
KOLLER: Yeah.

00:14:55.760 --> 00:14:59.520
LEE: So how does machine learning 
... how does AI help you?

00:14:59.520 --> 00:15:10.960
KOLLER: So I think when you look at how people 
currently select targets, it's a combination of  

00:15:10.960 --> 00:15:16.160
oftentimes at this point, with an increasing 
respect for the power of human genetics,  

00:15:16.160 --> 00:15:23.840
some search for a genetic association, oftentimes 
with a human-defined, highly subjective,  

00:15:23.840 --> 00:15:28.800
highly noisy clinical outcome, like some ICD 
[International Classification of Diseases] code.

00:15:28.800 --> 00:15:34.080
And those are often underpowered and very 
difficult to deconvolute the underlying  

00:15:34.080 --> 00:15:40.800
biology. You combine that with some mechanistic 
interrogation in a highly reductionist model  

00:15:40.800 --> 00:15:48.560
system looking at a small number of readouts, 
biochemical readouts, that a biologist thinks  

00:15:48.560 --> 00:15:56.320
are relevant to the disease. Like does this make 
this, whatever, cholesterol go up or amyloid  

00:15:56.320 --> 00:16:03.280
beta go down? Or whatever. And then you take 
that as the second stage, and you pick, based  

00:16:03.280 --> 00:16:08.240
on typically human intuition about, Oh, this one 
looks good to me, and then you take that forward.

00:16:08.240 --> 00:16:13.840
What we're doing is an attempt to be as unbiased 
and holistic as possible. So, first of all,  

00:16:13.840 --> 00:16:20.400
rather than rely on human-defined clinical 
endpoints, like this person has been diagnosed  

00:16:20.400 --> 00:16:27.520
with diabetes or fatty liver, we try and measure 
as much as we can a holistic physiological state  

00:16:27.520 --> 00:16:34.800
and then use machine learning to find structure, 
patterns in that human physiological readouts,  

00:16:34.800 --> 00:16:40.160
imaging readouts, and omics readouts from blood, 
from tissue, different kinds of imaging, and say,  

00:16:40.160 --> 00:16:44.960
these are different vectors that this 
disease takes, this group of individuals,  

00:16:44.960 --> 00:16:49.040
and here's a different group of individuals 
that maybe from a diagnostical perspective are  

00:16:49.040 --> 00:16:55.200
all called the same thing, but they are actually 
exhibiting a very different biology underlying it.

00:16:55.200 --> 00:17:02.160
And so that is something that doesn't emerge when 
a human being takes a reductionist view to looking  

00:17:02.160 --> 00:17:06.868
at this high-content data, and oftentimes, they 
don't even look at it and produce an ICD code.

00:17:06.868 --> 00:17:07.520
LEE: Right. Yep.

00:17:07.520 --> 00:17:12.000
KOLLER: The same approach, 
actually even the same code base,  

00:17:12.000 --> 00:17:18.320
is taken in the cellular data. So we don't 
just say, “Well, the thing that matters is,  

00:17:18.320 --> 00:17:23.760
you know, the total amount of lipid in 
the cell or whatever.” Rather, we say,  

00:17:23.760 --> 00:17:29.520
“Let's look at multiple readouts, multiple ways 
of looking at the cells, combine them using the  

00:17:29.520 --> 00:17:33.600
power of machine learning.” And again, looking 
at imaging readouts where a human's eyes just  

00:17:33.600 --> 00:17:39.120
glaze over looking at even a few dozen cells, 
far less a few hundreds of millions of cells,  

00:17:40.000 --> 00:17:46.240
and understand what are the different biological 
processes that are going on. What are the vectors  

00:17:46.240 --> 00:17:50.880
that the disease might take you in this direction, 
in this group of cells, or in that direction?

00:17:50.880 --> 00:17:58.080
And then importantly, we take all of that 
information from the human side, from the  

00:17:58.080 --> 00:18:03.360
cellular side, across these different readouts, 
and we combine them using an integrative approach  

00:18:03.360 --> 00:18:10.000
that looks at the combined weight of evidence 
and says, these are the targets that I have the  

00:18:10.000 --> 00:18:15.760
greatest amount of conviction about by looking 
across all of that information. Whereas we know,  

00:18:15.760 --> 00:18:21.040
and we know this, I'm sure you've seen this 
analysis done for clinicians, a human being  

00:18:21.040 --> 00:18:24.220
typically is able to keep three or four 
things in their head at the same time.

00:18:24.220 --> 00:18:26.560
LEE: Right.
KOLLER: A really good human being who's really

00:18:26.560 --> 00:18:29.360
expert at what they do can 
maybe get to six to eight.

00:18:29.360 --> 00:18:29.780
LEE: Yeah.

00:18:29.780 --> 00:18:32.400
KOLLER: The machine learning has 
no problem doing a few hundred.

00:18:32.400 --> 00:18:32.549
LEE: Right.

00:18:32.549 --> 00:18:37.920
KOLLER: And so you put that together, and 
that allows you, to your earlier question,  

00:18:37.920 --> 00:18:42.560
really select the targets around which 
you have the highest conviction. And  

00:18:42.560 --> 00:18:48.320
then those are the ones that we then 
prioritize for interrogation in more  

00:18:48.320 --> 00:18:54.320
expensive systems like mice and monkeys 
and then at the end of the day pick the  

00:18:54.320 --> 00:18:57.840
small handful that one can afford to 
actually take into clinical trials.

00:18:57.840 --> 00:19:01.760
LEE: So now, Insitro recently received $25 million  

00:19:01.760 --> 00:19:05.520
in milestone payments from Bristol 
Myers Squibb after discovering and  

00:19:05.520 --> 00:19:10.960
selecting a novel drug target for ALS. Can 
you tell us a little bit more about that?   

00:19:10.960 --> 00:19:16.720
KOLLER: We are incredibly excited 
about the first novel target,  

00:19:16.720 --> 00:19:22.640
and there is a couple of others just behind it 
in line that seem, you know, quite efficacious,  

00:19:22.640 --> 00:19:31.520
as well, that truly seem to reverse, albeit in a 
cellular system, what we now understand to be ALS  

00:19:31.520 --> 00:19:41.120
pathology across multiple different dimensions. 
There's been obviously many attempts made to try  

00:19:41.120 --> 00:19:46.000
and address ALS, which by the way, horrible, 
horrible disease, worse than most cancers.  

00:19:46.000 --> 00:19:50.640
It kills you almost inevitably in three to 
five years in a particularly horrific way.

00:19:51.760 --> 00:20:02.960
And what we have in our hands is a target that 
seems to revert a lot of the pathologies that  

00:20:02.960 --> 00:20:08.160
are associated with the disease, which we now 
understand has to do with the mis-splicing of  

00:20:08.160 --> 00:20:12.560
multiple proteins within the cell and 
creating defective versions of those  

00:20:12.560 --> 00:20:18.240
proteins that are just not operational. And 
we are seeing reversion of many of those.

00:20:18.880 --> 00:20:24.480
So can I tell you for sure it'll work in a 
human? No, there's many steps between now and  

00:20:24.480 --> 00:20:30.880
then. But we couldn't be more excited about 
the opportunity to provide what we hope will  

00:20:30.880 --> 00:20:36.680
be a disease-modifying intervention for these 
patients who really desperately need something.

00:20:36.680 --> 00:20:42.376
LEE: Well, it's certainly been making 
waves in the biotech and biomedical world.

00:20:42.376 --> 00:20:45.520
KOLLER: Thank you.
LEE: So we'll be really watching very closely.

00:20:46.160 --> 00:20:51.280
So, you know, I think just reflecting on, you 
know, what we missed and what we got right in  

00:20:51.280 --> 00:20:57.920
our book, I think in our book, we did have the 
insight that there would be an ability to connect,  

00:20:57.920 --> 00:21:05.200
say, genotypic and phenotypic data and, you know, 
just broadly the kinds of clinical measurements  

00:21:05.200 --> 00:21:11.520
that get made on real patients and that these 
things could be brought together. And I think  

00:21:11.520 --> 00:21:17.600
the work that you're doing really illustrates that 
in a very, very sophisticated, very ambitious way.

00:21:17.600 --> 00:21:24.000
But the fact that this could be connected all 
the way down to the biology, to the biochemistry,  

00:21:24.800 --> 00:21:30.525
I think we didn't have any clue what 
would happen, at least not this quickly.

00:21:30.525 --> 00:21:30.555
KOLLER: Well, I think the ...

00:21:30.555 --> 00:21:35.592
LEE: And I realize, you've been at this for 
quite a few years, but still, it's quite amazing.

00:21:35.592 --> 00:21:40.240
KOLLER: The thread that connects them is human 
genetics. And I think that has, to us, been,  

00:21:40.240 --> 00:21:45.600
sort of, the, kind of, the connective tissue 
that allows you to translate across different  

00:21:45.600 --> 00:21:50.880
systems and say, “What does this gene do? 
What does this gene do in this organ and  

00:21:50.880 --> 00:21:54.800
in that organ? What does it do in this 
type of cell and in that type of cell?”

00:21:54.800 --> 00:22:02.880
And then use that as sort of the thread, if you 
will, that follows the impact of modulating this  

00:22:02.880 --> 00:22:08.400
gene all the way from the simple systems where 
you can do the experiment to the complex systems  

00:22:08.400 --> 00:22:13.360
where you can't do the experiment until the 
very end, but you have the human genetics as  

00:22:13.360 --> 00:22:20.160
a way of looking at the statistics and 
understanding what the impact might be.

00:22:20.160 --> 00:22:29.520
LEE: So I'd like to now switch gears 
and take … I want to take two steps  

00:22:29.520 --> 00:22:35.280
in the remainder of this conversation towards 
the future. So one step into that future,  

00:22:35.280 --> 00:22:42.320
of course, we're living through now, which 
is just all of the crazy pace of work and  

00:22:42.320 --> 00:22:49.520
advancement in generative AI generally, 
you know, just the scale of transformers,  

00:22:49.520 --> 00:22:57.840
of post-training, and now inference scale 
and reasoning models and so on. And where  

00:22:57.840 --> 00:23:07.440
do you see all of that going with respect to 
the goals that you have and that Insitro has?

00:23:07.440 --> 00:23:15.760
KOLLER: So I think first and foremost is the 
parallel, if you will, to the predictions that  

00:23:15.760 --> 00:23:23.440
you focused on in your book, which is this will 
transform a lot of the core data processing tasks,  

00:23:23.440 --> 00:23:29.120
the information tasks. And sure, the doctors 
and nurses is one thing. But if you just think  

00:23:29.120 --> 00:23:35.200
of clinical trial operations or the submission 
of regulatory documents, these are all kind of  

00:23:36.000 --> 00:23:41.440
simple data … they're not simple, obviously, 
but they're data processing tasks. They involve  

00:23:41.440 --> 00:23:45.680
natural language. That's not going to 
be our focus, but I hope that others  

00:23:45.680 --> 00:23:52.400
will use that to make clinical trials 
faster, more efficient, less expensive.

00:23:52.400 --> 00:23:58.320
There's already a lot of progress that's 
happening on the molecular design side  

00:23:58.320 --> 00:24:03.920
of things and taking hypotheses and turning them 
quickly and effectively into molecules. As I said,  

00:24:03.920 --> 00:24:09.120
this is part of our work that we absolutely do 
and we don't talk about it very much, simply  

00:24:09.120 --> 00:24:15.920
because it's a very crowded landscape and a lot 
of companies are engaged on that. But I think it's  

00:24:15.920 --> 00:24:21.120
really important to be able to take biological 
insights and turn them into new molecules.

00:24:21.120 --> 00:24:28.400
And then, of course, the transformer models and 
their likes play a very significant role in that  

00:24:28.400 --> 00:24:34.560
sort of turning insights into molecules because 
you can have foundation models for proteins.  

00:24:34.560 --> 00:24:40.800
There are increasing efforts to create foundation 
models for other categories of molecules. And so  

00:24:40.800 --> 00:24:47.920
that will undoubtedly accelerate the process 
by which you can quickly generate different  

00:24:48.480 --> 00:24:54.216
molecular hypotheses and test them and learn from 
what you did so that you can do fewer iterations …

00:24:54.216 --> 00:24:54.228
LEE: Right.

00:24:54.228 --> 00:24:56.720
KOLLER: … before you converge 
on a successful molecule.

00:24:57.520 --> 00:25:03.360
I do think that arguably the biggest impact 
as yet to be had is in that understanding of  

00:25:03.360 --> 00:25:08.800
core human biology and what are the right 
ways to intervene in it. And that plays a  

00:25:08.800 --> 00:25:14.480
role in a couple different ways. First of 
all, it certainly plays a role in which  

00:25:14.480 --> 00:25:19.840
… if we are able to understand the 
human physiological state and, you know,  

00:25:19.840 --> 00:25:25.360
the state of different systems all the way 
down to the cell level, that will inform our  

00:25:25.360 --> 00:25:34.480
ability to pick hypotheses that are more likely 
to actually impact the right biologies underneath.

00:25:34.480 --> 00:25:34.533
LEE: Yep. Yeah.

00:25:34.533 --> 00:25:38.800
KOLLER: And the more data we're able to 
collect about humans and about cells,  

00:25:38.800 --> 00:25:44.080
the more successful our models will be at 
representing that human physiological state or  

00:25:44.080 --> 00:25:51.840
the cell biological state and making predictions 
reliably on the impact of these interventions.

00:25:51.840 --> 00:25:54.400
The other side of it, though, 
and this comes back, I think,  

00:25:54.400 --> 00:26:01.760
to themes that were very much in your book, is 
this will impact not only the early stages of  

00:26:01.760 --> 00:26:04.560
which hypotheses we interrogate, 
which molecules we move forward,  

00:26:04.560 --> 00:26:09.580
but also hopefully at the end of the day, 
which molecule we prescribe to which patient.

00:26:09.580 --> 00:26:09.592
LEE: Right.

00:26:09.592 --> 00:26:15.200
KOLLER: And I think there's been obviously so much 
narrative over the years about precision medicine,  

00:26:15.200 --> 00:26:20.000
personalized medicine, and very little of that 
has come to fruition, with the exception of,  

00:26:20.000 --> 00:26:24.800
you know, certain islands in oncology, 
primarily on genetically driven cancers.

00:26:26.080 --> 00:26:32.400
But I think the opportunity is still there. 
We just haven't been able to bring it to life  

00:26:32.400 --> 00:26:40.160
because of the lack of the right kind of data. 
And I think with the increasing amount of human,  

00:26:40.160 --> 00:26:45.840
kind of, foundational data that we're able 
to acquire, things that are not sort of  

00:26:46.640 --> 00:26:49.018
distilled through the eye of 
a clinician, for example, …

00:26:49.018 --> 00:26:49.028
LEE: Yes.

00:26:49.028 --> 00:26:53.120
KOLLER: … but really measurements of 
human pathology, we can start to get  

00:26:53.120 --> 00:26:59.760
to some of that precision, carving out 
of the human population and then get to  

00:26:59.760 --> 00:27:05.120
a world where we can prescribe the right 
medicine to the right patient and not only  

00:27:05.120 --> 00:27:11.120
in cancer but also in other diseases 
that are also not a single disease.

00:27:11.120 --> 00:27:17.680
LEE: All right, so now to wrap up this 
time together, I always try to ask one  

00:27:17.680 --> 00:27:22.960
more provocative last question. One 
of the dreams that comes naturally to  

00:27:22.960 --> 00:27:29.360
someone like me or any of my colleagues, 
probably even to you, is this idea of,  

00:27:29.360 --> 00:27:35.920
you know, wouldn't it be possible someday to 
have a foundation model for biology or for  

00:27:35.920 --> 00:27:41.760
human biology or foundation model for the 
human cell or something along these lines?

00:27:41.760 --> 00:27:44.880
And in fact, there are, of course, 
you and I are both aware of people  

00:27:44.880 --> 00:27:49.520
who are taking that idea seriously and 
chasing after it. I have people in our  

00:27:49.520 --> 00:27:55.560
labs that think hard about this kind of 
thing. Is it a reasonable thought at all?

00:27:55.560 --> 00:28:05.200
KOLLER: I have learned over the years to avoid 
saying the word never because technology proceeds  

00:28:05.200 --> 00:28:10.880
in ways that you often don't expect. And so 
will we at some point be able to measure the  

00:28:10.880 --> 00:28:16.160
cell in enough different ways across enough 
different channels at the same time that you  

00:28:16.160 --> 00:28:25.680
can piece together what a cell does? I think that 
is eminently feasible, not today, but over time.

00:28:27.200 --> 00:28:31.440
I don't think it's feasible using 
today's technology, although the  

00:28:31.440 --> 00:28:37.360
efforts to get there may expose where the 
biggest opportunities lie to, you know,  

00:28:37.360 --> 00:28:42.080
build that next layer. So I think it's good that 
people are working on really hard problems. I  

00:28:42.080 --> 00:28:47.680
would also point out that even if one were 
to solve that really challenging problem of  

00:28:47.680 --> 00:28:54.880
creating a model of a cell, there is thousands of 
different types of cells within the human body.

00:28:55.680 --> 00:28:59.338
They're very different. They 
also talk to each other …

00:28:59.338 --> 00:28:59.348
LEE: Yep.

00:28:59.348 --> 00:29:02.480
KOLLER: … both within the cell type 
and across different cell types. So  

00:29:02.480 --> 00:29:06.800
the combinatorial complexity 
of that system is, I think,  

00:29:06.800 --> 00:29:11.520
unfathomable to many people. I 
mean, I would say to all of us.

00:29:11.520 --> 00:29:12.080
LEE: Yeah.

00:29:12.080 --> 00:29:18.320
KOLLER: And so even from that very lofty goal, 
there is multiple big steps that would need to be  

00:29:18.320 --> 00:29:30.640
taken to a mechanistic model of the full organism. 
So will we ever get there? Again, you know,  

00:29:30.640 --> 00:29:39.520
I don't see a reason why this is impossible to 
do. So I think over time, technology will get  

00:29:39.520 --> 00:29:45.760
better and will allow us to build more and more 
elaborate models of more and more complex systems.

00:29:45.760 --> 00:29:47.410
Patients can't wait …

00:29:47.410 --> 00:29:47.428
LEE: Right. Yeah.

00:29:47.428 --> 00:29:51.760
KOLLER: … for that to happen in order 
for us to get them better medicines. So  

00:29:51.760 --> 00:29:57.360
I think there is a great basic science 
initiative on that side of things. And,  

00:29:57.360 --> 00:30:02.320
in parallel, we need to make do with 
the data that we have or can collect  

00:30:02.320 --> 00:30:08.960
or can print. We print a lot of data in our 
internal wet labs and get to drugs that are  

00:30:08.960 --> 00:30:13.000
effective even though they don't benefit 
from having a full-blown mechanistic model.

00:30:13.000 --> 00:30:18.720
LEE: Last question: where do you 
think we'll be in five years?

00:30:18.720 --> 00:30:24.080
KOLLER: Phew. If I had answered that question 
five years ago, I would have been very badly  

00:30:24.080 --> 00:30:31.840
embarrassed at the inaccuracy of my answer. 
[LAUGHTER] So I will not answer it today either.

00:30:32.400 --> 00:30:39.440
I will say that the thing about exponential 
curves is that they are very, very tricky,  

00:30:39.440 --> 00:30:48.320
and they move in unexpected ways. 
I would hope that in five years,  

00:30:48.320 --> 00:30:56.080
we will have made a sufficient investment in 
the generation of scientific data that we will  

00:30:56.080 --> 00:31:05.040
be able to move beyond data that was generated 
entirely by humans and therefore insights that  

00:31:05.040 --> 00:31:11.600
are derivative of what people already know 
to things that are truly novel discoveries.

00:31:11.600 --> 00:31:17.680
And I think in order to do that in, you know, 
math, maybe because math is entirely conceptual,  

00:31:17.680 --> 00:31:22.640
maybe you can do that today. Math is effectively 
a construct of the human mind. I don't think  

00:31:22.640 --> 00:31:26.640
biology is a construct of the human mind, 
and therefore one needs to collect enough  

00:31:26.640 --> 00:31:31.360
data to really build those models that 
will give rise to those novel insights.

00:31:31.360 --> 00:31:34.520
And that's where I hope we will have 
made considerable progress in five years.

00:31:34.520 --> 00:31:39.360
LEE: Well, I'm with you. I hope so, 
too. Well, you know, thank you, Daphne,  

00:31:39.360 --> 00:31:46.000
so much for this conversation. I learn a lot 
talking to you, and it was great to, you know,  

00:31:46.000 --> 00:31:51.000
connect again on this. And congratulations on 
all of this success. It's really groundbreaking.

00:31:51.000 --> 00:31:57.840
KOLLER: Thank you very much, Peter. It 
was a pleasure chatting with you, as well.

00:31:57.840 --> 00:32:00.943
[TRANSITION MUSIC]

00:32:00.943 --> 00:32:07.520
LEE: I still think of Daphne first and 
foremost as an AI researcher. And for sure,  

00:32:07.520 --> 00:32:11.440
her research work in machine learning 
continues to be incredibly influential to  

00:32:11.440 --> 00:32:17.840
this day. But it's her work on AI-enhanced drug 
development that now is on the verge of making  

00:32:17.840 --> 00:32:24.000
a really big difference on some of the most 
difficult diseases afflicting people today.

00:32:24.000 --> 00:32:27.920
In our book, Carey, Zak, and 
I predicted that AI might be  

00:32:27.920 --> 00:32:30.880
a meaningful accelerant in biomedical research,  

00:32:30.880 --> 00:32:37.200
but I don't know that we foresaw the incredible 
potential specifically in drug development.

00:32:37.200 --> 00:32:41.680
Today, we're seeing a flurry of activity 
at companies, universities, and startups  

00:32:41.680 --> 00:32:48.560
on generative AI systems that aid and maybe even 
completely automate the design of new molecules  

00:32:48.560 --> 00:32:56.000
as drug candidates. But now, in our conversation 
with Daphne, seeing AI go even further than that  

00:32:56.000 --> 00:33:01.760
to do what one might reasonably have assumed 
to be impossible, to identify and select novel  

00:33:01.760 --> 00:33:10.000
drug targets, especially for a neurodegenerative 
disease like ALS, it's just, well, mind blowing. 

00:33:10.000 --> 00:33:13.200
Let's continue our deep dive 
on AI and biomedical research  

00:33:13.200 --> 00:33:17.520
with this conversation with Noubar Afeyan:

00:33:21.320 --> 00:33:27.265
LEE: Noubar, thanks so much for joining. I'm 
really looking forward to this conversation.

00:33:27.265 --> 00:33:29.120
NOUBAR AFEYAN: Peter, thanks. Thrilled to be here.

00:33:29.120 --> 00:33:35.120
LEE: While I think most of the listeners to 
this podcast have heard of Flagship Pioneering,  

00:33:35.120 --> 00:33:42.160
it's still worth hearing from you, you know, 
what is Flagship? And maybe a little bit about  

00:33:42.160 --> 00:33:53.280
your background. And finally, you found a way to 
balance science and business creation. And so,  

00:33:53.280 --> 00:33:56.320
you know, your approach and 
philosophy to all of that.

00:33:56.320 --> 00:34:02.320
AFEYAN: Well, great. So maybe I'll just start 
out by way of quick background. You know,  

00:34:02.320 --> 00:34:06.400
my ... and since we're going talk 
about AI, I'll also highlight my  

00:34:06.400 --> 00:34:14.000
first contact with the topic of AI. So as an 
undergraduate in 1980 up at McGill University,  

00:34:14.000 --> 00:34:18.800
I was an engineering student, but I 
was really captivated by, at that time,  

00:34:18.800 --> 00:34:26.220
the talk on the campus around the expert system, 
heuristic-based, rule-based kind of programs.

00:34:26.220 --> 00:34:26.232
LEE: Right.

00:34:26.232 --> 00:34:31.440
AFEYAN: And so actually I had the dubious 
distinction of writing my one and only  

00:34:31.440 --> 00:34:37.760
college newspaper article. [LAUGHTER] That 
was a short career. And it was all about how  

00:34:37.760 --> 00:34:42.560
artificial intelligence would be impacting 
medicine, would be impacting, you know,  

00:34:42.560 --> 00:34:46.880
speech capture, translation, and some of the 
ideas that were there that it's interesting  

00:34:46.880 --> 00:34:53.680
to see now 45 years later re-emerge with 
some of the new learning-based models.

00:34:53.680 --> 00:35:00.320
My journey after college ended up taking 
me into biotechnology. In the early ’80s,  

00:35:00.320 --> 00:35:04.800
I came to MIT to do a PhD. At the time, the 
field was brand new. I ended up being the  

00:35:04.800 --> 00:35:11.920
first PhD graduate from MIT in this combination 
biology and engineering degree. And since then,  

00:35:11.920 --> 00:35:17.200
I've basically been—so since 1987—a 
founder, a technologist in the space  

00:35:17.200 --> 00:35:21.520
of biotechnology for human health 
and as well for planetary health.

00:35:21.520 --> 00:35:27.200
And then in 1999/2000 formed what is now 
Flagship Pioneering, which essentially  

00:35:27.200 --> 00:35:33.040
was an attempt to bring together the three 
elements of what we know are important in  

00:35:33.040 --> 00:35:39.040
startups. That is scientific capital, human 
capital, and financial capital. Right now,  

00:35:39.040 --> 00:35:43.600
startups get that from different places. The 
science in our fields mostly come from academia,  

00:35:43.600 --> 00:35:47.176
research hospitals. The human 
capital comes from other startups …

00:35:47.176 --> 00:35:47.188
LEE: Yeah.

00:35:47.188 --> 00:35:51.280
AFEYAN: … or large companies or some academics 
leave. And then the financial capital is usually  

00:35:51.280 --> 00:35:55.680
venture capital, but there's also now more 
and more other deeper pockets of money.

00:35:55.680 --> 00:36:01.760
What we thought was, what if all that existed in 
one entity and instead of having to convince each  

00:36:01.760 --> 00:36:07.440
other how much they should believe the other 
if we just said, “Let's use that power to go  

00:36:07.440 --> 00:36:12.480
work on much further out things”? But in a way 
where nobody would believe it in the beginning,  

00:36:12.480 --> 00:36:16.800
but we could give ourselves a little 
bit of time to do impactful big things.

00:36:16.800 --> 00:36:20.160
Twenty-five years later, that's 
the road we've stayed on.

00:36:20.160 --> 00:36:27.600
LEE: OK. So let's get into AI. Now, you 
know, what I've been asking guests is  

00:36:27.600 --> 00:36:33.600
kind of an origin story. And there's the 
origin story of contact with AI, you know,  

00:36:33.600 --> 00:36:37.920
before the emergence of generative AI and 
afterwards. I don't think there's much of  

00:36:37.920 --> 00:36:48.400
a point to asking you the pre-ChatGPT. But … 
so let's focus on your first encounter with  

00:36:48.400 --> 00:36:54.720
ChatGPT or generative AI. When did that 
happen, and what went through your head?

00:36:54.720 --> 00:36:58.080
AFEYAN: Yeah. So, if you permit 
me, Peter, just for very briefly,  

00:36:58.080 --> 00:37:02.960
let me actually say I had the 
interesting opportunity over the  

00:37:02.960 --> 00:37:07.491
last 25 years to actually stay pretty 
close to the machine learning world …

00:37:07.491 --> 00:37:07.508
LEE: Yeah. Yeah.

00:37:07.508 --> 00:37:12.000
AFEYAN: … because one, as you well know, 
among the most prolific users of machine  

00:37:12.000 --> 00:37:17.840
learning has been the bioinformatics computational 
biology world because it's been so data rich that  

00:37:17.840 --> 00:37:22.720
anything that can be done, people have thrown at 
these problems because unlike most other things,  

00:37:22.720 --> 00:37:27.200
we're not working on man-made data. We're 
looking at data that comes from nature,  

00:37:27.200 --> 00:37:30.720
the complexity of which far 
exceeds our ability to comprehend.

00:37:30.720 --> 00:37:36.400
So you could imagine that any approach 
to statistically reduce complexity,  

00:37:36.400 --> 00:37:41.520
get signal out of scant data—that's 
a problem that's been around.

00:37:41.520 --> 00:37:45.120
The other place where I've been exposed to this, 
which I'm going to come back to because that's  

00:37:45.120 --> 00:37:50.240
where it first felt totally different to me, 
is that some 25 years ago, actually the very  

00:37:50.240 --> 00:37:55.840
first company we started was a company that 
attempted to use evolutionary algorithms to  

00:37:55.840 --> 00:38:02.960
essentially iteratively evolve consumer-packaged 
goods online. Literally, we tried to, you know,  

00:38:02.960 --> 00:38:10.240
consider features of products as genes and create 
little genomes of them. And by recombination and  

00:38:10.240 --> 00:38:14.720
mutation, we could create variety. And 
then we could get people through panels  

00:38:14.720 --> 00:38:20.320
online—this was 2002/2003 timeframe—we could 
essentially get people through iterative cycles  

00:38:20.320 --> 00:38:24.640
of voting to create a survival of the fittest. 
And that's a company that was called Affinnova.

00:38:24.640 --> 00:38:29.280
The reason I say that is that I knew 
that there’s a much better way to do  

00:38:29.280 --> 00:38:33.897
this if only: one, you can generate variety …

00:38:33.897 --> 00:38:33.908
LEE: Yeah.

00:38:33.908 --> 00:38:38.480
AFEYAN: … without having to prespecify 
genes. We couldn’t do that before. And,  

00:38:38.480 --> 00:38:43.360
two, which we’ve come back to nowadays, 
you can actually mimic how humans think  

00:38:43.360 --> 00:38:46.560
about voting on things and just 
get rid of that element of it.

00:38:46.560 --> 00:38:52.080
So then to your question of when does this kind of 
begin to feel different? So you could imagine that  

00:38:52.080 --> 00:38:57.760
in biotechnology, you know, as an engineer 
by background, I always wanted to do CAD,  

00:38:57.760 --> 00:39:02.080
and I picked the one field in which CAD doesn't 
exist, which is biology. Computer-aided design  

00:39:02.080 --> 00:39:07.577
is kind of a notional thing in that space. 
But boy, have we tried. For a long time, …

00:39:07.577 --> 00:39:07.588
LEE: Yep.

00:39:07.588 --> 00:39:11.680
AFEYAN: … people would try to do, you know, 
hidden Markov models of genomes to try to figure  

00:39:11.680 --> 00:39:16.240
out what should be the next, you know, base 
that you may want to or where genes might be,  

00:39:16.240 --> 00:39:23.360
etc. But the notion of generating in biology has 
been something we've tried for a while. And in the  

00:39:23.360 --> 00:39:30.000
late teens, so kind of 2018, ’17, ’18, because 
we saw deep learning come along, and you could  

00:39:30.000 --> 00:39:35.280
basically generate novelty with some of the deep 
learning models … and so we started asking, “Could  

00:39:35.280 --> 00:39:42.640
you generate a protein basically by training a 
correspondence table, if you will, between protein  

00:39:42.640 --> 00:39:48.060
structures and their underlying DNA sequence?” Not 
their protein sequence, but their DNA sequence.

00:39:48.060 --> 00:39:50.880
LEE: Yeah.
AFEYAN: So that's a big leap. So ’17/’18,

00:39:50.880 --> 00:39:56.480
we started this thing. It was called 56. It 
was FL56, Flagship Labs 56, our 56th project.

00:39:56.480 --> 00:40:01.280
By the way, we started this parallel one called 
“57” that did it in a very different way. So  

00:40:01.280 --> 00:40:06.560
one of them did pure black box model-building. The 
other one said, you know what, we don't want to do  

00:40:06.560 --> 00:40:12.160
the kind of ... at that time, AlphaFold was in its 
very early embodiments. And we said, “Is there a  

00:40:12.160 --> 00:40:18.520
way we could actually take little, you know, multi 
amino acid kind of almost grammars, if you will,  

00:40:18.520 --> 00:40:22.880
a little piece, and then see if we could compose 
a protein that way?” So we were experimenting.

00:40:22.880 --> 00:40:28.400
And what we found was that actually, if you 
show enough instances and you could train a  

00:40:28.400 --> 00:40:32.320
transformer model—back in the day, that's 
what we were using—you could actually,  

00:40:32.320 --> 00:40:37.360
say, predict another sequence that should 
have the same activity as the first one.

00:40:37.360 --> 00:40:37.453
LEE: Yeah.

00:40:37.453 --> 00:40:39.680
AFEYAN: So we trained on green 
fluorescent proteins. Now,  

00:40:39.680 --> 00:40:44.800
we're talking about seven years ago. We trained 
on enzymes, and then we got to antibodies.

00:40:44.800 --> 00:40:49.120
With antibodies, we started seeing that, boy, 
this could be a pretty big deal because it has  

00:40:49.120 --> 00:40:54.480
big market impact. And we started bringing 
in some of the diffusion models that were  

00:40:54.480 --> 00:40:58.640
beginning to come along at that time. And so 
we started getting much more excited. This was  

00:40:58.640 --> 00:41:03.985
all done in a company that subsequently got 
renamed from FL56 to Generate:Biomedicines, …

00:41:03.985 --> 00:41:04.000
LEE: Yep, yep.

00:41:04.000 --> 00:41:08.480
AFEYAN: … which is one of the leaders 
in protein design using the generative  

00:41:08.480 --> 00:41:12.000
techniques. It was interesting because 
Generate:Biomedicines is a company that  

00:41:12.000 --> 00:41:17.920
was called that before generative AI was a 
thing, [LAUGHTER] which was kind of very ironic.

00:41:17.920 --> 00:41:24.240
And, of course, that team, which operates 
today very, very kind of at the cutting edge,  

00:41:24.240 --> 00:41:29.200
has published their models. They came up with this 
first Chroma model, which is a diffusion-based  

00:41:29.200 --> 00:41:34.000
model, and then started incorporating a lot 
of the LLM capabilities and fusing them.

00:41:34.000 --> 00:41:38.000
Now we're doing atomistic models and 
many other things. The point being,  

00:41:38.000 --> 00:41:42.371
that gave us a glimpse of how 
quickly the capability was gaining, …

00:41:42.371 --> 00:41:42.388
LEE: Yeah. Yeah.

00:41:42.388 --> 00:41:46.000
AFEYAN: … just like evolution shows you. 
Sometimes evolution is super silent,  

00:41:46.000 --> 00:41:49.120
and then all of a sudden, all hell 
breaks loose. And that's what we saw.

00:41:49.120 --> 00:41:54.400
LEE: Right. One of the things that 
I reflect on just in my own journey  

00:41:54.400 --> 00:42:00.080
through this is there are other emotions 
that come up. One that was prominent for  

00:42:00.080 --> 00:42:07.040
me early on was skepticism. Were there 
points when even in your own work,  

00:42:07.040 --> 00:42:14.560
transformer-based work on this early on, that you 
had doubts or skepticism that these transformer  

00:42:14.560 --> 00:42:19.160
architectures would be or diffusion-based 
approaches would be worth anything?

00:42:19.160 --> 00:42:23.920
AFEYAN: You know, it's interesting, 
I think that, I'm going to say this  

00:42:23.920 --> 00:42:28.240
to you in a kind of a friendly way, but you'll 
understand what I mean. In the world I live in,  

00:42:28.240 --> 00:42:32.240
it's kind of like the slums of innovation, 
[LAUGHTER] kind of like just doing things  

00:42:32.240 --> 00:42:38.080
that are not supposed to work. The 
notion of skepticism is a luxury,  

00:42:38.080 --> 00:42:43.280
right. I assume everything we do won't 
work. And then once in a while I'm wrong.

00:42:43.280 --> 00:42:48.880
And so I don't actually try to evaluate 
whether before I bring something in,  

00:42:48.880 --> 00:42:53.840
like just think about it. We, 
some hundred or so times a year,  

00:42:53.840 --> 00:43:00.320
ask “what if” questions that lead us to totally 
weird places of thought. We then try to iterate,  

00:43:00.320 --> 00:43:04.800
iterate, iterate to come up with something that's 
testable. Then we go into a lab, and we test it.

00:43:04.800 --> 00:43:07.600
So in that world, right, 
sitting there going, like,  

00:43:07.600 --> 00:43:11.440
“How do I know this transformer is going 
to work?” The answer is, “For what?” Like,  

00:43:11.440 --> 00:43:16.720
it's going to work. To make something up ... 
well, guess what? We knew early on with LLMs  

00:43:16.720 --> 00:43:20.560
that hallucination was a feature, 
not a bug for what we wanted to do.

00:43:20.560 --> 00:43:27.840
So it's just such a different use that, of 
course, I have trained scientific skepticism,  

00:43:27.840 --> 00:43:34.400
but it's a little bit like looking at a 
competitive situation in an ecology and saying,  

00:43:34.400 --> 00:43:39.520
“I bet that thing's going to die.” Well, you'd be 
right—most of the time, you'd be right. [LAUGHTER]

00:43:39.520 --> 00:43:46.560
So I just don't … like, it … and that's 
why—I guess, call me an early adopter—for us,  

00:43:46.560 --> 00:43:52.635
things that could move the needle even a little, 
but then upon repetition a lot, let alone this, …

00:43:52.635 --> 00:43:53.188
LEE: Yeah.

00:43:53.188 --> 00:43:55.920
AFEYAN: … you have to embrace. 
You can't wait there and say,  

00:43:55.920 --> 00:43:59.200
I'll embrace it once it's ready. 
And so that's what we did.

00:43:59.200 --> 00:44:08.000
LEE: Hmm. All right. So let's get into some 
specifics and what you are seeing either in your  

00:44:08.000 --> 00:44:15.680
portfolio companies or in the research projects 
or out in the industry. What is going on today  

00:44:15.680 --> 00:44:24.240
with respect to AI really being used for something 
meaningful in the design and development of drugs?

00:44:24.240 --> 00:44:29.280
AFEYAN: In companies that are doing as diverse 
things as—let me give you a few examples—a  

00:44:30.720 --> 00:44:35.680
project that's now become a named company 
called ProFound Therapeutics that literally  

00:44:35.680 --> 00:44:39.840
discovered three, four years ago, and would 
not have been able to without some of the  

00:44:39.840 --> 00:44:47.440
big data-model-building capabilities, 
that our cells make literally thousands,  

00:44:47.440 --> 00:44:53.040
if not tens of thousands, of more 
proteins than we were aware of, full stop.

00:44:53.040 --> 00:44:57.338
We had done the human genome sequence, there 
was 20,000 genes, we thought that there was …

00:44:57.338 --> 00:45:00.320
LEE: Wow.
AFEYAN: … maybe 70-80,000, 100,000 proteins,

00:45:00.320 --> 00:45:05.120
and that's that. And it turns out that our 
cells have a penchant to express themselves  

00:45:05.120 --> 00:45:10.320
in the form of proteins, and they have 
many other ways than we knew to do that.

00:45:10.320 --> 00:45:16.160
Now, so what does that mean? That means that 
we have generated a massive amount of data,  

00:45:16.160 --> 00:45:20.800
the interpretation of which, the use 
of which to guide what you do and what  

00:45:20.800 --> 00:45:25.120
these things might be involved with 
is purely being done using the most  

00:45:25.120 --> 00:45:29.737
cutting-edge data-trained models that 
allow you to navigate such complexity.

00:45:29.737 --> 00:45:29.752
LEE: Wow. Hmm.

00:45:29.752 --> 00:45:33.840
AFEYAN: That's just one example. Another 
example: a company called Quotient Therapeutics,  

00:45:33.840 --> 00:45:36.000
again three, four years old. I can 
talk about the ones that are three,  

00:45:36.000 --> 00:45:39.280
four years old because we've kind of 
gotten to a place where we've decided  

00:45:39.280 --> 00:45:43.120
that it's not going to fail yet, 
[LAUGHTER] so we can talk about it.

00:45:43.120 --> 00:45:48.000
You know, we discovered—our team 
discovered—that in our cells, right,  

00:45:48.000 --> 00:45:51.600
so we know that when we get cancer, 
our cells have genetic mutations in  

00:45:51.600 --> 00:45:57.280
them or DNA mutations that are correlated and 
often causal to the hyperproliferative stages  

00:45:57.280 --> 00:46:02.640
of cancer. But what we assume is that all 
the other cells in our body, pretty much,  

00:46:02.640 --> 00:46:07.920
have one copy of their genes from our mom, 
one copy from our dad, and that's that.

00:46:07.920 --> 00:46:11.440
And when very precise deep sequencing came along,  

00:46:11.440 --> 00:46:14.540
we always asked the question, “How 
much variation is there cell to cell?”

00:46:14.540 --> 00:46:14.552
LEE: Right.

00:46:14.552 --> 00:46:19.600
AFEYAN: And the answer was it's kind of noise, 
random variation. Well, our team said, “Well,  

00:46:19.600 --> 00:46:25.600
what if it's not really that random?” because upon 
cell division cycles, there's selection happening  

00:46:25.600 --> 00:46:31.954
on these cells. And so not just in cancer but 
in liver cells, in muscle cells, in skin cells …

00:46:31.954 --> 00:46:32.139
LEE: Oh, interesting.

00:46:32.139 --> 00:46:37.680
AFEYAN: … can you imagine that there's an 
evolutionary experiment that is favoring either  

00:46:38.240 --> 00:46:43.840
compensatory mutations that are helping you 
avoid disease or disease-caused mutations that  

00:46:43.840 --> 00:46:49.840
are gaining advantage as a way to understand the 
mechanism? Sure enough—I wouldn't be telling you  

00:46:49.840 --> 00:46:55.920
otherwise—with massive amount of single cell 
sequencing from individual patient samples,  

00:46:55.920 --> 00:47:02.240
we've now discovered that the human genome is 
mutated on average in our bodies 10,000 times,  

00:47:02.240 --> 00:47:05.360
like over every base, like, it's huge numbers.

00:47:05.360 --> 00:47:10.320
And we're finding very interesting 
big signals come out of this massive  

00:47:10.320 --> 00:47:15.280
amount of data. By the way, data 
of the sort that the human mind,  

00:47:15.280 --> 00:47:19.335
if it tries to assign causal 
explanations to what's happening …

00:47:19.335 --> 00:47:20.760
LEE: Right.
AFEYAN: … is completely inadequate.

00:47:20.760 --> 00:47:24.880
LEE: When you think about a language 
model, we're learning from human language,  

00:47:24.880 --> 00:47:31.360
and the totality of human language—at least 
relative to what we're able to compute today  

00:47:31.360 --> 00:47:37.200
in terms of constructing a model—the totality 
of human language is actually pretty limited.  

00:47:37.200 --> 00:47:41.200
And in fact, you know, as is always 
written about in click-baity titles,  

00:47:41.200 --> 00:47:44.880
you know, the big model builders 
are actually starting to run short.

00:47:44.880 --> 00:47:46.920
AFEYAN: Running out, running out, yes. [LAUGHTER]

00:47:46.920 --> 00:47:51.680
LEE: But one of the things that perplexes 
me and maybe even worries me—like these  

00:47:51.680 --> 00:48:00.320
two examples—are generally in the realm 
of cellular biology and the complexity.  

00:48:00.320 --> 00:48:06.800
Let's just take the example of your company, 
ProFound. You know, the complexity of what's  

00:48:06.800 --> 00:48:15.840
going on and the potential genetic diversity 
is such that, can we ever have enough data?  

00:48:15.840 --> 00:48:22.232
You know, because there just aren't that many 
human beings. There just aren't that many samples.

00:48:22.232 --> 00:48:25.760
AFEYAN: Well, it depends 
on what you want to train,  

00:48:25.760 --> 00:48:30.720
right. So if you want to train a de novo 
evolutionary model that could take you  

00:48:30.720 --> 00:48:36.400
from bacteria to human mammalian cells and 
the like, there may not be—and I'm not an  

00:48:36.400 --> 00:48:40.080
expert in that—but that's a question 
that we often kind of think about.

00:48:41.040 --> 00:48:47.920
But if you're trying to train a ... like 
you know what the proteins we know about,  

00:48:47.920 --> 00:48:52.400
how they interact with pathways and disease 
mechanisms and the like. Now all of a sudden  

00:48:52.400 --> 00:48:58.080
you find out that there's a whole continent of 
them missing in your explanations. But there are  

00:48:58.080 --> 00:49:04.000
things you can reason, in quotations, through 
analogy, functional analogy, sequence analogy,  

00:49:04.000 --> 00:49:08.880
homology. So there's a lot of things that 
we could do to essentially make use of this,  

00:49:08.880 --> 00:49:14.000
even though you may not have 
the totality of data needed to,  

00:49:14.000 --> 00:49:18.240
kind of, predict, based on a de novo 
sequence, exactly what it's going to do.

00:49:18.240 --> 00:49:24.720
So I agree with the comparison. But ... 
but you're right. The complexity is … just  

00:49:24.720 --> 00:49:28.960
keep in mind, on average, a protein may be 
interacting with 50 to 100 other proteins.

00:49:28.960 --> 00:49:29.056
LEE: Right.

00:49:29.056 --> 00:49:32.480
AFEYAN: So if you find thousands 
of proteins, you've found a massive  

00:49:32.480 --> 00:49:38.400
interaction space through which information 
is being processed in a living cell.

00:49:38.400 --> 00:49:43.600
LEE: But do you find in your 
AI companies that access to  

00:49:43.600 --> 00:49:49.360
data ends up being a key challenge? 
Or, you know, how central is that?

00:49:49.360 --> 00:49:54.960
AFEYAN: Access to data is a key challenge for 
the companies we have that are trying to build  

00:49:54.960 --> 00:50:00.800
just models. But that's the minority of things we 
do. The majority of things we do is to actually  

00:50:00.800 --> 00:50:06.240
co-develop the data and the models. And as 
you know well, because you guys, you know,  

00:50:06.240 --> 00:50:12.160
have given us some ideas around this space, that, 
you know, you could generate data and then think  

00:50:12.160 --> 00:50:15.973
about what you're to do with it, which is the 
way biotech is operated with bioinformatics.

00:50:15.973 --> 00:50:15.992
LEE: Right, right.

00:50:15.992 --> 00:50:20.800
AFEYAN: Or you could generate bespoke data 
that is used to train the model that's quite  

00:50:20.800 --> 00:50:25.440
separate from what you would have done in the 
natural course of biology. So we're doing much  

00:50:25.440 --> 00:50:31.840
more of the latter of late, and I think that'll 
continue. So, but these things are proliferating.

00:50:31.840 --> 00:50:39.920
I mean, it's hard to find a place where we're 
not using this. And the “this” is any and all  

00:50:39.920 --> 00:50:47.840
data-driven model building, generative, LLM-based, 
but also every other technique to make progress.

00:50:47.840 --> 00:50:54.480
LEE: Sure. So now moving away from the 
straight biochemistry applications,  

00:50:54.480 --> 00:51:01.760
what about AI in the process of building a 
business, of making investment decisions,  

00:51:01.760 --> 00:51:06.240
of actually running an operation? 
What are you seeing there?

00:51:06.240 --> 00:51:11.920
AFEYAN: So, well, you know, Moderna, which is a 
company that I'm quite proud of being a founder  

00:51:11.920 --> 00:51:19.600
and chairman of, has adopted a significant, 
significant amount of AI embedded into their  

00:51:19.600 --> 00:51:27.040
operations in all aspects: from the manufacturing, 
quality control, the clinical monitoring,  

00:51:27.040 --> 00:51:32.160
the design—every aspect. And in fact, they've had 
a partnership that they've had for a little while  

00:51:32.160 --> 00:51:36.800
here with OpenAI, and they've tried many different 
ways to stay at the cutting edge of that.

00:51:36.800 --> 00:51:39.840
So we see that play out at 
some scale. That’s a 5,000-,  

00:51:39.840 --> 00:51:44.160
6,000-person organization, and what 
they're doing is a good example of  

00:51:44.160 --> 00:51:49.040
what early adopters would do, at least 
in our kind of biotechnology company.

00:51:49.040 --> 00:51:56.320
But then, you know, in our space, I would say 
the efficiency impact is kind of no different,  

00:51:56.320 --> 00:52:02.320
than, you know, anywhere else in academia you 
might adopt it or in other kinds of companies.  

00:52:02.320 --> 00:52:07.280
But where I find it an interesting kind of 
maybe segue is the degree to which it may  

00:52:07.280 --> 00:52:13.899
fundamentally change the way we think about how 
to do science, which is a whole other use, right?

00:52:13.899 --> 00:52:13.912
LEE: Right.

00:52:13.912 --> 00:52:18.480
AFEYAN: So it's not an efficiency gain per 
se, although it's maybe an effectiveness  

00:52:18.480 --> 00:52:26.709
gain when it comes to science, but can you just 
fundamentally train models to generate hypotheses?

00:52:26.709 --> 00:52:26.720
LEE: Yep.

00:52:26.720 --> 00:52:30.240
AFEYAN: And we have done that, and we've 
been doing this for the last three years.  

00:52:30.240 --> 00:52:33.360
And now it's getting better and better, the 
better these reasoning engines are getting  

00:52:33.360 --> 00:52:40.400
and kind of being able to extrapolate and 
train for novelty. Can you convert that to  

00:52:40.400 --> 00:52:46.720
the world's best experimental protocol to very 
precisely falsify your hypothesis, on and on?

00:52:46.720 --> 00:52:50.160
That closing of that loop, kind of 
what we call autonomous science,  

00:52:50.160 --> 00:52:54.160
which we've been trying to do for the last two, 
three years and are making some progress in,  

00:52:54.160 --> 00:52:59.040
that to me is another kind of 
bespoke use of these things,  

00:52:59.040 --> 00:53:04.640
not to generate molecules in its chemistry, but 
to change the behavior of how science is done.

00:53:04.640 --> 00:53:10.720
LEE: Yeah. So I always end with a 
couple of provocative questions,  

00:53:10.720 --> 00:53:19.760
but I need—before we do that, while we're on 
this subject—to get your take on Lila Sciences.

00:53:19.760 --> 00:53:24.080
And there is a vision there that I think is very  

00:53:24.080 --> 00:53:27.600
interesting. It'd be great 
to hear it described by you.

00:53:27.600 --> 00:53:32.000
AFEYAN: Sure. So Lila, after operating for 
two to three years in kind of a preparatory  

00:53:32.000 --> 00:53:37.440
kind of stealth mode, we've now had a 
little bit more visibility around, and  

00:53:37.440 --> 00:53:42.640
essentially what we're trying to do there is to 
create what we call automated science factories,  

00:53:42.640 --> 00:53:49.360
and such a factory would essentially be able to 
take problems, either computationally specified  

00:53:49.360 --> 00:53:57.040
or human-specified, and essentially do 
the experimental work in order to either  

00:53:57.680 --> 00:54:02.000
make an optimization happen or enable 
something that just didn’t exist.  

00:54:02.880 --> 00:54:08.462
And it’s really, at this point, we’ve 
shown proof of concept in narrow areas.

00:54:08.462 --> 00:54:08.472
LEE: Yep.

00:54:08.472 --> 00:54:10.880
AFEYAN: But it’s hard to 
say that if you can do this,  

00:54:10.880 --> 00:54:15.040
you can’t do some other things, so we’re 
just expanding it that way. We don’t think  

00:54:15.040 --> 00:54:20.067
we need a complete proof or complete 
demonstration of it for every aspect.

00:54:20.067 --> 00:54:20.080
LEE: Right.

00:54:20.080 --> 00:54:24.080
AFEYAN: So we're just kind of being 
opportunistic. The idea for Lila is  

00:54:24.080 --> 00:54:27.920
to partner with a number of companies. 
The good news is, within Flagship,  

00:54:27.920 --> 00:54:31.360
there's 48 of them. And so there's a whole 
lot of them they can partner with to get  

00:54:31.360 --> 00:54:36.480
their learning cycles. But eventually 
they want to be a real alternative to  

00:54:36.480 --> 00:54:42.400
every time somebody has an idea, having to 
kind of go into a lab and manually do this.

00:54:42.400 --> 00:54:46.298
I do want to say one thing we touched on, 
Peter, though, just on that front, which is ...

00:54:46.298 --> 00:54:46.308
LEE: Yep.

00:54:46.308 --> 00:54:49.600
AFEYAN: ... if you say, like, “What 
problem is this going to solve?” It's  

00:54:49.600 --> 00:54:55.440
several but an important one is 
just the flat-out human capacity  

00:54:55.440 --> 00:55:01.520
to reason on this much data and 
this much complexity that is real.  

00:55:01.520 --> 00:55:07.520
Because nature doesn't try to abstract 
itself in a human understandable form.

00:55:07.520 --> 00:55:07.760
LEE: Right. Yeah.

00:55:07.760 --> 00:55:14.000
AFEYAN: In biology, since it's kind of like 
progress happens through evolutionary kind  

00:55:14.000 --> 00:55:20.480
of selections, the evidence of which [has] 
long been lost, and so therefore, you just  

00:55:20.480 --> 00:55:26.480
see what you have, and then it has a behavior. I 
really do think that there's something to be said,  

00:55:26.480 --> 00:55:30.720
and I want to—just for your audience—lay out 
a provocative, at least, thought on all this,  

00:55:30.720 --> 00:55:36.160
which Lila is a beginning embodiment of, which is 
that I really think that what's going to happen  

00:55:36.160 --> 00:55:43.760
over the next five, 10 years, even while we're 
all fascinated with the impending arrival of AGI  

00:55:43.760 --> 00:55:48.000
[artificial general intelligence] is really 
what I call poly-intelligence, which is the  

00:55:48.000 --> 00:55:55.200
combination of human intelligence, machine 
intelligence, AI, and nature's intelligence.

00:55:55.200 --> 00:56:01.120
We're all fascinated at the human-machine 
interface. We know the human-nature interface,  

00:56:01.120 --> 00:56:04.400
but imagine the machine-nature interface—that is,  

00:56:04.400 --> 00:56:10.560
actually letting loose a digital 
kind of information processing life  

00:56:10.560 --> 00:56:17.040
form through the algorithms that are being 
developed and the commensurately complex,  

00:56:17.040 --> 00:56:22.560
maybe much more complex. We'll see. And so now 
the question becomes, what does the human do?

00:56:22.560 --> 00:56:27.280
And we're living in a world which is human 
dominated, which means the humans say, “If I don't  

00:56:27.280 --> 00:56:32.480
understand it, it's not real, basically. And if 
I don't understand it, I can't regulate it.” And  

00:56:32.480 --> 00:56:37.680
we're going to have to make peace with the fact 
that we're not going to be able to predictably  

00:56:37.680 --> 00:56:42.960
affect things without necessarily understanding 
them the way we could if we just forced ourselves  

00:56:42.960 --> 00:56:47.760
to only work on problems we can understand. 
And that world we're not ready for at all.

00:56:47.760 --> 00:56:54.320
LEE: Yeah. All right. So this one I predict is 
going to be a little harder for you because I  

00:56:54.320 --> 00:57:00.880
think while you think about the future, you 
live very much in the present. But I'd like  

00:57:00.880 --> 00:57:09.200
you to make some predictions about what the 
biotech and biopharmaceutical industries are  

00:57:09.200 --> 00:57:14.480
going to be able to do two years from now, 
five years from now, 10 years from now.

00:57:14.480 --> 00:57:18.080
AFEYAN: Yeah, well, it's hard for 
me because you know my nature,  

00:57:18.080 --> 00:57:19.840
which is that I think this is all emergent.

00:57:19.840 --> 00:57:20.360
LEE: Right.

00:57:20.360 --> 00:57:28.800
AFEYAN: And so I would be the conceit of 
predicting. So I would say with likelihood  

00:57:28.800 --> 00:57:34.800
positive predictive value of less than 10%, 
I'm happy to answer your question. So I'm  

00:57:34.800 --> 00:57:42.000
not trying to score high [LAUGHTER] because 
I really think that my job is to envision it,  

00:57:42.000 --> 00:57:44.069
not to predict it. And that's 
a little bit different, right?

00:57:44.069 --> 00:57:47.360
LEE: Yeah, I actually was trying to 
pick what would be the hardest possible  

00:57:47.360 --> 00:57:49.800
question I could ask you, [LAUGHTER] 
and this is what I came up with.

00:57:49.800 --> 00:57:59.920
AFEYAN: Yeah, no, no, I'm kidding here. So now 
look, I think that we will cross this threshold of  

00:58:01.360 --> 00:58:06.800
understandability. And of course you're seeing 
that in a lot of LLM things today. And of course,  

00:58:06.800 --> 00:58:10.560
people are trying to train for things 
that are explainers and all that whole,  

00:58:10.560 --> 00:58:14.480
there's a whole world of that. But I 
think at some point we're going to have  

00:58:14.480 --> 00:58:19.280
to kind of let go and get comfortable 
working on things that, you know …

00:58:19.280 --> 00:58:24.320
I sometimes tell people, you know, and I'm not the 
first, but scientists and engineers are different,  

00:58:24.320 --> 00:58:27.600
it's said, in that engineers work 
on things that they don't wait  

00:58:27.600 --> 00:58:30.800
until they get a full understanding 
of before they work with them. Well,  

00:58:30.800 --> 00:58:33.735
now scientists are going to have 
to get used to that, too, right?

00:58:33.735 --> 00:58:33.752
LEE: Yeah. Yeah.

00:58:33.752 --> 00:58:39.920
AFEYAN: Because insisting that it's only valid 
if it's understandable. So, I would say, look,  

00:58:39.920 --> 00:58:47.680
I hope that the time … for example, I think major 
improvements will be made in patient selection.  

00:58:47.680 --> 00:58:53.898
If we can test drugs on patients that are more 
synchronized as to the stage of their disease …

00:58:53.898 --> 00:58:53.908
LEE: Yep.

00:58:53.908 --> 00:58:56.960
AFEYAN: ... I think the answer will be much 
better. We're working on that. It's a company  

00:58:56.960 --> 00:59:02.080
called Etiome, very, very early stage. It's 
really beautiful data, very early data that  

00:59:02.080 --> 00:59:04.640
shows that when we talk about MASH [metabolic 
dysfunction-associated steatohepatitis], liver  

00:59:04.640 --> 00:59:10.720
disease, when we talk about Parkinson's, there's 
such a heterogeneity, not only of the subset type  

00:59:10.720 --> 00:59:15.120
of the disease, but the stage of the disease, 
that this notion that you have stage one cancer,  

00:59:15.120 --> 00:59:20.960
stage two cancer, again, nobody told nature 
there's stages of that kind. It's a continuum.

00:59:20.960 --> 00:59:28.480
But if you can synchronize based on training, kind 
of, the ability to detect who are the patients  

00:59:28.480 --> 00:59:34.320
that are in enough of a close proximity that 
should be treated so that the trial—much smaller a  

00:59:34.320 --> 00:59:39.520
trial size—could give you a drug, then afterwards, 
you can prescribe it using these approaches.

00:59:40.400 --> 00:59:43.680
Kind of we're going to find that what 
we thought is one disease is more like  

00:59:43.680 --> 00:59:49.280
15 diseases. That's bad news because we're 
not going to be able to claim that we can  

00:59:49.280 --> 00:59:52.240
treat everything which we can. It's good 
news in that there's going to be people  

00:59:52.240 --> 00:59:55.500
who are going to start making much 
more specific solutions to things.

00:59:55.500 --> 00:59:55.512
LEE: Right.

00:59:55.512 --> 01:00:03.840
AFEYAN: So I can imagine that. I can imagine a 
generation of, kind of, students who are going  

01:00:03.840 --> 01:00:11.040
to be able to play in this space without having 
25 years of graduate education on the subject.  

01:00:11.040 --> 01:00:16.160
So what is deemed knowledge sufficient 
to do creative things will change. I can  

01:00:16.160 --> 01:00:21.080
go on and on, but I think all this is 
very close by and it's very exciting.

01:00:21.080 --> 01:00:24.240
LEE: Noubar, I just always have so much fun,  

01:00:24.240 --> 01:00:29.600
and I learn really a lot. It's high-density 
learning when I talk to you. And so  

01:00:29.600 --> 01:00:35.200
I hope our listeners feel the same way. 
It's something I really appreciate.

01:00:35.200 --> 01:00:38.640
AFEYAN: Well, Peter, thanks for this. 
And I think your listeners know that if  

01:00:38.640 --> 01:00:42.800
I was asking you questions, you would 
be answering them with equal if not  

01:00:42.800 --> 01:00:53.576
more fascinating stuff. So, thanks for 
giving me the chance to do that today.

01:00:53.576 --> 01:00:53.595
[TRANSITION MUSIC]

01:00:53.595 --> 01:00:57.280
LEE: I’m always fascinated by Noubar’s 
perspectives on fundamental research  

01:00:57.280 --> 01:01:01.520
and how it connects to human health and 
the building of successful companies.  

01:01:01.520 --> 01:01:07.600
I see him as a classic “systems thinker,” and 
by that, I mean he builds impressive things like  

01:01:07.600 --> 01:01:13.600
Flagship Pioneering itself, which he created 
as a kind of biomedical innovation system.

01:01:13.600 --> 01:01:17.280
In our conversation, I was really 
struck by the fact that he’s been  

01:01:17.280 --> 01:01:21.120
thinking about the potential impact 
of transformers—transformers being  

01:01:21.120 --> 01:01:26.320
the fundamental building block of large 
language models—as far back as 2017,  

01:01:26.320 --> 01:01:32.080
when the first paper on the attention mechanism 
in transformers was published by Google.

01:01:32.080 --> 01:01:37.360
But, you know, it isn’t only about using AI to 
do things like understand and design molecules  

01:01:37.360 --> 01:01:44.240
and antibodies faster. It's interesting that he is 
also pushing really hard towards a future where AI  

01:01:44.240 --> 01:01:51.840
might “close the loop” from hypothesis generation, 
to experiment design, to analysis, and so on.

01:01:51.840 --> 01:01:57.880
Now, here’s my conversation with Dr. Eric Topol:

01:01:57.880 --> 01:02:03.348
LEE: Eric, it's really great to have you here.

01:02:03.348 --> 01:02:07.200
ERIC TOPOL: Oh, Peter, I'm thrilled 
to be here with you here at Microsoft.

01:02:07.200 --> 01:02:11.280
LEE: You're a super famous person. 
Extremely well known to researchers  

01:02:11.280 --> 01:02:16.800
even in computer science, as we 
have here at Microsoft Research.

01:02:16.800 --> 01:02:25.584
But the question I'd like to ask is, how would 
you explain to your parents what you do every day?

01:02:25.584 --> 01:02:28.960
TOPOL: [LAUGHS] That's a good question. If I was  

01:02:28.960 --> 01:02:33.120
just telling them I'm trying to come up 
with better ways to keep people healthy,  

01:02:33.120 --> 01:02:38.320
that probably would be the easiest way to do 
it because if I ever got in deeper, I would  

01:02:38.320 --> 01:02:45.501
lose them real quickly. They're not around, but 
just thinking about what they could understand.

01:02:45.501 --> 01:02:45.513
LEE: Right.

01:02:45.513 --> 01:02:50.960
TOPOL: I think as long as they knew 
it was work centered on innovative  

01:02:50.960 --> 01:02:55.800
paths to promoting and preserving human 
health, that would get to them, I think.

01:02:55.800 --> 01:03:02.400
LEE: OK, so now, kind of the second topic, 
and then we let the conversation flow,  

01:03:03.520 --> 01:03:08.560
is about origin stories with respect 
to AI. And with most of our guests,  

01:03:08.560 --> 01:03:16.240
you know, I factor that into two pieces: 
the encounters with AI before ChatGPT and  

01:03:16.240 --> 01:03:20.480
what we call generative AI and 
then the first contacts after.

01:03:20.480 --> 01:03:28.480
And, of course, you have extensive contact 
with both now. But let's start with  

01:03:29.120 --> 01:03:36.360
how you got interested in machine learning 
and AI prior to ChatGPT. How did that happen?

01:03:36.360 --> 01:03:46.240
TOPOL: Yeah, it was out of necessity. So back, you 
know, when I started at Scripps at the end of ’06,  

01:03:46.240 --> 01:03:51.040
we started accumulating, you know, massive 
datasets. First, it was whole genomes.  

01:03:52.080 --> 01:03:56.880
We did one of the early big cohorts of 1,400  

01:03:56.880 --> 01:04:01.200
people of healthy aging. We called 
the Wellderly whole genome sequence.

01:04:01.200 --> 01:04:04.160
And then we started big in the sensor world,  

01:04:04.160 --> 01:04:06.720
and then we started saying, what are 
we going to do with all this data,  

01:04:06.720 --> 01:04:12.160
with electronic health records and all 
those sensors? And now we got whole genomes.

01:04:12.160 --> 01:04:13.920
And basically, what we were doing,  

01:04:13.920 --> 01:04:18.640
we were in hoarding mode. We didn't 
have a way to meaningfully analyze it.

01:04:18.640 --> 01:04:19.360
LEE: Right.

01:04:19.360 --> 01:04:23.360
TOPOL: You would read about how, 
you know, data is the new oil and,  

01:04:23.360 --> 01:04:29.200
you know, gold and whatnot. But we just 
didn't have a way to extract the juice.  

01:04:30.960 --> 01:04:37.040
And even when we wanted to analyze 
genomes, it was incredibly laborious.

01:04:37.040 --> 01:04:38.160
LEE: Yeah.

01:04:38.160 --> 01:04:43.280
TOPOL: And we weren't extracting a 
lot of the important information.  

01:04:43.280 --> 01:04:48.720
So that's why … not having any 
training in computer science,  

01:04:48.720 --> 01:04:56.560
when I was doing the ... about three years of work 
to do the book Deep Medicine, I started really,  

01:04:57.120 --> 01:05:03.120
first auto-didactic about, you know, machine 
learning. And then I started contacting a  

01:05:03.120 --> 01:05:09.920
lot of the real top people in the field and 
hanging out with them, and learning from them,  

01:05:09.920 --> 01:05:15.280
getting their views as to, you know, where we 
are today, what models are coming in the future.

01:05:15.280 --> 01:05:20.480
And then I said, “You know what? We are going to 
be able to fix this mess.” [LAUGHS] We're going  

01:05:20.480 --> 01:05:25.920
to get out of the hoarding phase, and we're going 
to get into, you know, really making a difference.

01:05:25.920 --> 01:05:30.320
So that's when I embraced the 
future of AI. And I knew, you know,  

01:05:30.320 --> 01:05:34.960
back—that was six years ago when it was 
published and probably eight or nine  

01:05:34.960 --> 01:05:40.240
years ago when I was doing the research, 
and I knew that we weren't there yet.

01:05:40.240 --> 01:05:44.880
You know, at the time, we were seeing the 
image interpretation. That was kind of the  

01:05:44.880 --> 01:05:53.360
early promise. But really, the models that 
were transformative, the transformer models,  

01:05:53.360 --> 01:05:58.148
they were incubating back in 2017. 
So people knew something was brewing.

01:05:58.148 --> 01:06:00.800
LEE: Right. Yes.
TOPOL: And everyone said we're going to get there.

01:06:00.800 --> 01:06:09.440
LEE: So then, ChatGPT comes out November 
of 2022; there’s GPT-4 in 2023, and now a  

01:06:09.440 --> 01:06:17.720
lot has happened. Do you remember what your 
first encounter with that technology was?

01:06:17.720 --> 01:06:27.120
TOPOL: Oh, sure. First, ChatGPT. You 
know, in the last days of November ’22,  

01:06:27.120 --> 01:06:33.840
I was just blown away. I mean, I'm having 
a conversation. I'm having fun. And this is  

01:06:33.840 --> 01:06:40.400
humanoid responding to me. I said, “What?” You 
know? So that was to me, a moment I'll never  

01:06:40.400 --> 01:06:49.920
forget. And so I knew that the world was, you 
know, at a very kind of momentous changing point.

01:06:49.920 --> 01:06:56.080
Of course, knowing, too, that this is going to 
be built on, and built on quickly. Of course,  

01:06:56.080 --> 01:07:01.360
I didn't know how soon GPT-4 and all 
the others were going to come forward,  

01:07:01.360 --> 01:07:12.480
but that was a wake-up call that the capabilities 
of AI had just made a humongous jump, which  

01:07:12.480 --> 01:07:17.111
seemingly was all of a sudden, although 
I did know this had been percolating …

01:07:17.111 --> 01:07:17.509
LEE: Right.

01:07:17.509 --> 01:07:22.960
TOPOL: … you know, for what, 
at least five years, that,  

01:07:22.960 --> 01:07:28.160
you know, it really was getting 
into its position to do this.

01:07:28.160 --> 01:07:33.760
LEE: I know one of the things that was challenging 
psychologically and emotionally for me is,  

01:07:33.760 --> 01:07:39.280
it made me rethink a lot of things that 
were going on in Microsoft Research in  

01:07:39.280 --> 01:07:46.400
areas like causal reasoning, natural language 
processing, speech processing, and so on.

01:07:46.400 --> 01:07:53.920
I'm imagining you must have had some emotional 
struggles too because you have this amazing book,  

01:07:53.920 --> 01:07:59.120
Deep Medicine. Did you have to … did it 
go through your mind to rethink what you  

01:07:59.120 --> 01:08:03.640
wrote in Deep Medicine in light of this 
or, or, you know, how did that feel?

01:08:03.640 --> 01:08:09.920
TOPOL: It's funny you ask that because 
in this one chapter I have on the virtual  

01:08:10.480 --> 01:08:13.915
health coach, I wrote a 
whole bunch of scenarios ...

01:08:13.915 --> 01:08:14.080
LEE: Yeah.

01:08:14.080 --> 01:08:21.200
TOPOL: … that were very kind of futuristic. 
You know, about how the AI interacts with the  

01:08:21.200 --> 01:08:26.240
person's health and schedules their appointment 
for this and their scan and tells them what lab  

01:08:26.240 --> 01:08:29.280
tests they should tell their doctor to 
have, and, you know, all these things.  

01:08:29.280 --> 01:08:33.280
And I sent a whole bunch of these, thinking 
that they were a little too far-fetched.

01:08:33.280 --> 01:08:33.960
LEE: Yes.

01:08:33.960 --> 01:08:37.200
TOPOL: And I sent them to my editor 
when I wrote the book, and he says,  

01:08:37.200 --> 01:08:41.680
“Oh, these are great. You should 
put them all in.” [LAUGHTER] What  

01:08:41.680 --> 01:08:46.440
I didn't realize is they weren't that, 
you know, they were all going to happen.

01:08:46.440 --> 01:08:47.913
LEE: Yeah. They weren't that far-fetched at all.

01:08:47.913 --> 01:08:51.600
TOPOL: Not at all. If there's one 
thing I've learned from all this,  

01:08:51.600 --> 01:08:54.520
is our imagination isn't big enough.

01:08:54.520 --> 01:08:55.000
LEE: Yeah.

01:08:55.000 --> 01:08:56.640
TOPOL: We think too small.

01:08:56.640 --> 01:09:04.480
LEE: Now in our book that Carey, Zak, and 
I wrote, you know, we made, you know, we  

01:09:04.480 --> 01:09:14.160
sort of guessed that GPT-4 might help biomedical 
researchers, but I don't think that any of us had  

01:09:14.160 --> 01:09:23.440
the thought in mind that the architecture around 
generative AI would be so directly applicable to,  

01:09:23.440 --> 01:09:32.240
you know, say, protein structures or, you 
know, to clinical health records and so on.

01:09:32.240 --> 01:09:36.960
And so a lot of that seems much more obvious 
today. But two years ago, it wasn't. But we  

01:09:36.960 --> 01:09:43.120
did guess that biomedical researchers would 
find this interesting and be helped along.

01:09:43.120 --> 01:09:46.160
So as you reflect over the past two years,  

01:09:46.160 --> 01:09:54.080
you know, do you have things that you 
think are very important, kind of,  

01:09:54.080 --> 01:09:59.480
meaningful applications of generative AI 
in the kinds of research that Scripps does?

01:09:59.480 --> 01:10:02.080
TOPOL: Yeah. I mean, I think for one,  

01:10:02.080 --> 01:10:05.280
you pointed out, I’ll never forget, about 
how the term generative AI is a misnomer.

01:10:05.280 --> 01:10:05.920
LEE: Yeah.

01:10:05.920 --> 01:10:15.280
TOPOL: And so it really was prescient about how, 
you know, it had a pluripotent capability in every  

01:10:15.280 --> 01:10:23.840
respect, you know, of editing and creating. So 
that was something that I think was telling us,  

01:10:25.040 --> 01:10:30.960
an indicator that this is, you know, a lot 
bigger than how it's being labeled. And our  

01:10:30.960 --> 01:10:41.040
expectations can actually be more than what we 
had seen previously with the earlier version.

01:10:41.040 --> 01:10:49.280
So I think what's happened is that now, 
we keep jumping. It's so quick that we  

01:10:49.280 --> 01:10:53.760
can't … you know, first we think, oh, 
well, we’ve gone into the agentic era,  

01:10:53.760 --> 01:10:58.221
and then we could pass that with reasoning. 
[LAUGHTER] And, you know, we just can't …

01:10:58.221 --> 01:10:59.240
LEE: Right.
TOPOL: It's just wild.

01:10:59.240 --> 01:10:59.760
LEE: Yeah.

01:10:59.760 --> 01:11:11.120
TOPOL: So I think so many of us now will put 
in prompts that will necessitate or ideally  

01:11:11.920 --> 01:11:22.640
result in a not-immediate gratification, 
but rather one that requires, you know,  

01:11:22.640 --> 01:11:28.395
quite a bit of combing through 
the corpus of knowledge ...

01:11:28.395 --> 01:11:28.720
LEE: Yeah.

01:11:28.720 --> 01:11:33.840
TOPOL: … and getting, with all the citations,  

01:11:33.840 --> 01:11:42.800
a report or a response. And I think now this 
has been a reset because to do that on our own,  

01:11:43.600 --> 01:11:48.080
it takes, you know, many, many 
hours. And it's usually incomplete.

01:11:48.800 --> 01:11:53.280
But one of the things that was so 
different in the beginning was you  

01:11:53.280 --> 01:11:58.720
would get the references from up 
to a year and a half previously.

01:11:58.720 --> 01:12:00.220
LEE: Yep.

01:12:00.220 --> 01:12:02.555
TOPOL: And that's not good enough. [LAUGHS]

01:12:02.555 --> 01:12:03.000
LEE: Right.

01:12:03.000 --> 01:12:08.070
TOPOL: And now you get references, 
like, from the day before.

01:12:08.070 --> 01:12:12.080
LEE: Yes. Yeah.
TOPOL: And so, you say, “Why would you do

01:12:12.080 --> 01:12:16.400
a regular search for anything when 
you could do something like this?”

01:12:16.400 --> 01:12:16.880
LEE: Yeah.

01:12:16.880 --> 01:12:22.480
TOPOL: And then, you know, the reasoning 
power. And a lot of people who are not  

01:12:22.480 --> 01:12:26.942
using this enough still are talking 
about, “Well, there's no reasoning.”

01:12:26.942 --> 01:12:26.953
LEE: Yeah.

01:12:26.953 --> 01:12:30.720
TOPOL: Which you dealt with really 
well in the book. But what, of course,  

01:12:30.720 --> 01:12:34.560
you couldn't have predicted is the new dimensions.

01:12:34.560 --> 01:12:35.320
LEE: Right.

01:12:35.320 --> 01:12:42.000
TOPOL: I think you nailed it with GPT-4. But it's 
all these just, kind of, stepwise progressions  

01:12:42.000 --> 01:12:49.040
that have been occurring because of the velocity 
that's unprecedented. I just can't believe it.

01:12:49.040 --> 01:12:54.640
LEE: We were aware of the idea of 
multi-modality, but we didn't appreciate,  

01:12:54.640 --> 01:12:59.040
you know, what that would mean. Like AlphaFold 
[protein structure database], you know,  

01:12:59.040 --> 01:13:05.360
the ability for AI to understand—or 
crystal structures—to really start  

01:13:05.360 --> 01:13:11.200
understanding something more fundamental 
about biochemistry or medicinal chemistry.

01:13:11.200 --> 01:13:15.200
I have to admit, when we wrote 
the book, we really had no idea.

01:13:15.200 --> 01:13:20.480
TOPOL: Well, I feel the same way. I still 
today can't get over it because the reason  

01:13:20.480 --> 01:13:23.840
AlphaFold and Demis [Hassabis] and John 
Jumper [AlphaFold’s co-creators] were  

01:13:23.840 --> 01:13:26.320
so successful is there was this protein databank.

01:13:26.320 --> 01:13:26.880
LEE: Yes.

01:13:26.880 --> 01:13:32.640
TOPOL: And it had been kept for decades. 
And so, they had the substrate to work with.

01:13:32.640 --> 01:13:33.240
LEE: Right.

01:13:33.240 --> 01:13:38.040
TOPOL: So, you say, “OK, we can do proteins.” 
But then how do you do everything else?

01:13:38.040 --> 01:13:38.400
LEE: Right.

01:13:38.400 --> 01:13:43.600
TOPOL: And so this whole, what I call, 
“large language of life model” work,  

01:13:43.600 --> 01:13:46.320
which has gone into high 
gear like I've never seen.

01:13:46.320 --> 01:13:46.800
LEE: Yeah.

01:13:46.800 --> 01:13:51.982
TOPOL: You know, now to this holy 
grail of a virtual cell, and ...

01:13:51.982 --> 01:13:51.993
LEE: Yeah.

01:13:51.993 --> 01:14:00.160
TOPOL: You know, it's basically ... it's ... it 
was inspired by proteins. But now it's hitting on,  

01:14:00.160 --> 01:14:07.160
you know, ligands and small molecules, cells. 
I mean, nothing is being held back here.

01:14:07.160 --> 01:14:07.680
LEE: Yeah.

01:14:07.680 --> 01:14:10.400
TOPOL: So how could anybody have predicted that?

01:14:10.400 --> 01:14:10.640
LEE: Right.

01:14:10.640 --> 01:14:13.840
TOPOL: I sure wouldn't have thought 
it would be possible at this point.

01:14:13.840 --> 01:14:18.320
LEE: Yeah. So just to challenge you, 
where do you think that is going to  

01:14:18.320 --> 01:14:25.040
be two years from now? Five years from now? 
Ten years from now? Like, so you talk about  

01:14:25.040 --> 01:14:31.200
a virtual cell. Is that achievable within 
10 years, or is that still too far out?

01:14:31.200 --> 01:14:35.440
TOPOL: No, I think within 10 years for sure. You  

01:14:35.440 --> 01:14:39.560
know the group that got assembled 
that Steve Quake pulled together?

01:14:39.560 --> 01:14:39.960
LEE: Right.

01:14:39.960 --> 01:14:45.440
TOPOL: I think has 42 authors in a paper in 
Cell. The fact that he could get these 42  

01:14:45.440 --> 01:14:51.355
experts in life science and some in computer 
science to come together and all agree …

01:14:51.355 --> 01:14:51.669
LEE: Yeah.

01:14:51.669 --> 01:14:53.280
TOPOL: … that not only is this a worthy goal,  

01:14:53.280 --> 01:14:58.000
but it's actually going to be 
realized, that was impressive.

01:14:58.000 --> 01:15:01.920
I challenged him about that. How did 
you get these people all to agree?  

01:15:02.960 --> 01:15:07.600
So many of them were naysayers. And by the 
time the workshop finished, they were fully  

01:15:07.600 --> 01:15:15.440
convinced. I think that what we're seeing is so 
much progress happening so quickly. And then all  

01:15:15.440 --> 01:15:21.901
the different models, you know, across DNA, 
RNA, and everything are just zooming forward.

01:15:21.901 --> 01:15:21.913
LEE: Yeah.

01:15:21.913 --> 01:15:25.360
TOPOL: And it's just a matter of pulling 
this together. Now when we have that,  

01:15:25.360 --> 01:15:30.240
and I think it could easily be well 
before a decade and possibly, you know,  

01:15:30.240 --> 01:15:38.240
between the five- and 10-year mark—that's just 
a guess—but then we're moving into another  

01:15:38.240 --> 01:15:45.200
era of life science because right now, you 
know, this whole buzz about drug discovery.

01:15:45.200 --> 01:15:46.400
LEE: Yep.

01:15:46.400 --> 01:15:51.680
TOPOL: It's not... with the ability to do 
all these perturbations at a cellular level.

01:15:51.680 --> 01:15:53.662
LEE: Right.
TOPOL: Or the cell of interest.

01:15:53.662 --> 01:15:53.673
LEE: Yeah.

01:15:53.673 --> 01:15:58.640
TOPOL: Or the cell-to-cell interactions or the 
intra-cell interaction. So once you nail that,  

01:15:58.640 --> 01:16:07.200
yeah, it takes it to a kind of another predictive 
level that we haven't really fathomed. So, yes,  

01:16:07.200 --> 01:16:10.480
there's going to be drug discovery 
that's accelerated. But this would  

01:16:10.480 --> 01:16:14.080
make that and also the underpinnings of diseases.

01:16:14.080 --> 01:16:14.800
LEE: Yeah.

01:16:14.800 --> 01:16:17.520
TOPOL: So the idea that there's so many diseases  

01:16:17.520 --> 01:16:22.235
we don't understand now. And 
if you had virtual cell, …

01:16:22.235 --> 01:16:24.955
LEE: Yeah.
TOPOL: … you would probably get to that answer …

01:16:24.955 --> 01:16:25.429
LEE: Yeah.

01:16:25.429 --> 01:16:29.440
TOPOL: … much more quickly. So 
whether it's underpinnings of diseases  

01:16:30.000 --> 01:16:32.960
or what it's going to take 
to really come up with far  

01:16:32.960 --> 01:16:37.200
better treatments—preventions—I think 
that's where virtual cell will get us.

01:16:37.200 --> 01:16:44.400
LEE: There's a technical question ... I wonder if 
you have an opinion. You may or may not. There is  

01:16:44.400 --> 01:16:50.720
sort of what I would refer to as ab initio 
approaches to this. You know, you start from  

01:16:50.720 --> 01:16:58.000
the fundamental physics and chemistry, and we know 
the laws, we have the math and, you know, we can  

01:16:58.000 --> 01:17:05.600
try to derive from there … in fact, we can even 
run simulations of that math to generate training  

01:17:05.600 --> 01:17:14.240
data to build generative models and work up to 
a cell, or forget all of that and just take as  

01:17:14.240 --> 01:17:24.000
many observations and measurements of, say, living 
cells as possible, and just have faith that hidden  

01:17:24.000 --> 01:17:29.920
amongst all of the observational data, there 
is structure and language that can be derived.

01:17:29.920 --> 01:17:33.840
So that's sort of bottom-up 
versus top-down approaches.  

01:17:33.840 --> 01:17:36.880
Do you have an opinion about which way?

01:17:36.880 --> 01:17:43.200
TOPOL: Oh, I think you go after both. And clearly 
whenever you're positing that you've got a virtual  

01:17:43.200 --> 01:17:49.280
cell model that's working, you've got to do 
the traditional methods as well to validate it,  

01:17:49.280 --> 01:17:54.960
and … so all that. You know, I think if 
you're going to go out after this seriously,  

01:17:54.960 --> 01:17:59.640
you have to pull out all the stops. Both 
approaches, I think, are going to be essential.

01:17:59.640 --> 01:18:07.120
LEE: You know, if what you're saying is true, 
and it is amazing to hear the confidence, the one  

01:18:07.120 --> 01:18:15.680
thing I tried to explain to someone nontechnical 
is that for a lot of problems in medicine, we  

01:18:15.680 --> 01:18:21.920
just don't have enough data in a really profound 
way. And the most profound way to say that is,  

01:18:21.920 --> 01:18:28.080
since Adam and Eve, there have only been an 
estimated 106 billion people who have ever lived.

01:18:28.080 --> 01:18:35.760
So even if we had the DNA of every human 
being, every individual of Homo sapiens,  

01:18:35.760 --> 01:18:39.760
there are certain problems for 
which we would not have enough data.

01:18:39.760 --> 01:18:40.240
TOPOL: Sure.

01:18:40.240 --> 01:18:43.840
LEE: And so I think another 
thing that seems profound to me,  

01:18:43.840 --> 01:18:50.059
if we can actually have a virtual cell, is 
we can actually make trillions of virtual …

01:18:50.059 --> 01:18:50.071
TOPOL: Yeah

01:18:50.071 --> 01:18:54.240
LEE: … human beings. The true genetic 
diversity could be realized for our species.

01:18:54.240 --> 01:19:00.640
TOPOL: I think you nailed it. The 
ability to have that type of data,  

01:19:00.640 --> 01:19:06.000
no less synthetic data, I 
mean, it’s just extraordinary.

01:19:06.000 --> 01:19:06.880
LEE: Yeah.

01:19:06.880 --> 01:19:12.000
TOPOL: We will get there someday. I'm confident 
of that. We may be wrong in projections. And  

01:19:12.000 --> 01:19:15.680
I do think [science writer] Philip Ball 
won't be right that it will never happen,  

01:19:15.680 --> 01:19:20.760
though. [LAUGHTER] No, I think that if 
there's a holy grail of biology, this is it.

01:19:20.760 --> 01:19:21.200
LEE: Yeah.

01:19:21.200 --> 01:19:28.160
TOPOL: And I think you're absolutely 
right about where that will get us.

01:19:28.160 --> 01:19:29.440
LEE: Yeah.

01:19:29.440 --> 01:19:33.120
TOPOL: Transcending the beginning of the species.

01:19:33.120 --> 01:19:33.670
LEE: Yeah.

01:19:33.670 --> 01:19:34.640
TOPOL: Of our species.

01:19:34.640 --> 01:19:41.680
LEE: Yeah. All right. So now, we're starting to 
run short on time here. And so I wanted to ask  

01:19:41.680 --> 01:19:47.680
you about, I'm in my 60s, so I actually think 
about this a lot more. [LAUGHTER] And I know  

01:19:47.680 --> 01:19:53.120
you've been thinking a lot about longevity. 
And, of course, your new book, Super Agers.

01:19:53.120 --> 01:19:58.080
And one of the reasons I'm so eager to read 
is it's a topic very top of mind for me and  

01:19:58.080 --> 01:20:03.440
actually for a lot of people. Where is this 
going? Because this is another area where you  

01:20:03.440 --> 01:20:08.953
hear so much hype. At the same time, 
you see Nobel laureate scientists ...

01:20:08.953 --> 01:20:09.431
TOPOL: Yeah.

01:20:09.431 --> 01:20:10.240
LEE: ... working on this.

01:20:10.240 --> 01:20:10.795
TOPOL: Yeah.

01:20:10.795 --> 01:20:13.520
LEE: So, so what's, what's real there?

01:20:13.520 --> 01:20:19.840
TOPOL: Yeah. Well, it's really … the real deal 
is the science of aging is zooming forward.

01:20:19.840 --> 01:20:29.920
And that's exciting. But I see it bifurcating. On 
the one hand, all these new ideas, strategies to  

01:20:29.920 --> 01:20:37.520
reverse aging are very ambitious. Like cell 
reprogramming and senolytics and, you know,  

01:20:38.400 --> 01:20:42.640
the rejuvenation of our thymus 
gland, and it's a long list.

01:20:42.640 --> 01:20:43.120
LEE: Yeah.

01:20:43.120 --> 01:20:45.680
TOPOL: And they’re really cool science,  

01:20:45.680 --> 01:20:53.520
and it used to be the mouse lived longer. 
Now it's the old mouse looks really young.

01:20:53.520 --> 01:20:54.393
LEE: Yeah. Yeah.

01:20:54.393 --> 01:20:58.800
TOPOL: All the different features. A blind 
mouse with cataracts is all of a sudden  

01:20:58.800 --> 01:21:07.200
there's no cataracts. I mean, so these things are 
exciting, but none of them are proven in people,  

01:21:07.200 --> 01:21:13.520
and they all have significant risk, no less, 
you know, the expense that might be attached.

01:21:13.520 --> 01:21:14.360
LEE: Right.

01:21:14.360 --> 01:21:17.840
TOPOL: And some people are jumping 
the gun. They're taking rapamycin,  

01:21:17.840 --> 01:21:22.480
which can really knock out their immune 
system. So they all carry a lot of risk.  

01:21:22.480 --> 01:21:28.080
And people are just getting a little 
carried away. We're not there yet.

01:21:28.080 --> 01:21:33.440
But the other side, which is what I 
emphasize in the book, which is exciting,  

01:21:33.440 --> 01:21:38.760
is that we have all these new metrics 
that came out of the science of aging.

01:21:38.760 --> 01:21:39.280
LEE: Yes.

01:21:39.280 --> 01:21:45.280
TOPOL: So we have clocks of the body. Our 
biological clock versus our chronological clock,  

01:21:45.280 --> 01:21:50.960
and we have organ clocks. So I can say, you 
know, Peter, we've assessed all your organs  

01:21:50.960 --> 01:21:57.680
and your immune system. And guess what? Every one 
of them is either at or less than your actual age.

01:21:57.680 --> 01:21:58.160
LEE: Right.

01:21:58.160 --> 01:22:03.360
TOPOL: And that's very reassuring. And by the 
way, your methylation clock is also … I don't  

01:22:03.360 --> 01:22:08.320
need to worry about you so much. And then 
I have these other tests that I can do now,  

01:22:08.320 --> 01:22:14.640
like, for example, the brain. We have 
an amazing protein p-Tau217 that we  

01:22:14.640 --> 01:22:19.738
can say over 20 years in advance 
of you developing Alzheimer's, …

01:22:19.738 --> 01:22:19.749
LEE: Yeah.

01:22:19.749 --> 01:22:27.360
TOPOL: … we can look at that, and it's modifiable 
by lifestyle, bringing it down. It should be you  

01:22:27.360 --> 01:22:35.440
can change the natural history. So what we've seen 
is an explosion of knowledge of metrics, proteins,  

01:22:35.440 --> 01:22:41.680
no less, you know, our understanding 
at the gene level, the gut microbiome,  

01:22:41.680 --> 01:22:47.680
the immune system. So that's what's so exciting. 
How our immune system ages. Immunosenescence. How  

01:22:47.680 --> 01:22:55.680
we have more inflammation—inflammaging—with aging. 
So basically, we have three diseases that kill us,  

01:22:55.680 --> 01:23:01.400
that take away our health: heart, 
cancer, and neurodegenerative.

01:23:01.400 --> 01:23:02.040
LEE: Yep.

01:23:02.040 --> 01:23:06.960
TOPOL: And they all take more than 
20 years. They all have a defective  

01:23:06.960 --> 01:23:12.320
immune system inflammation problem, and 
they're all going to be preventable.

01:23:12.320 --> 01:23:13.160
LEE: Yeah.

01:23:13.160 --> 01:23:19.590
TOPOL: That's what's so exciting. So we don't have 
to have reverse aging. We can actually work on …

01:23:19.590 --> 01:23:21.029
LEE: Just prevent aging in the first place.

01:23:21.029 --> 01:23:26.080
TOPOL: … the age-related diseases. So basically, 
what it means is: I got to find out if you have a  

01:23:26.080 --> 01:23:31.440
risk, if you're in this high-risk group for this 
particular condition, because if you are—and we  

01:23:31.440 --> 01:23:38.000
have many levels, layers, orthogonal ways 
to check—we don't just bank it all on one  

01:23:38.000 --> 01:23:43.760
polygenic test. We're going to have several 
ways, say this is the one we are going ...

01:23:43.760 --> 01:23:50.640
And then we go into high surveillance, where, 
let's say if it's your brain, we do more p-Tau, if  

01:23:50.640 --> 01:23:56.640
we need to do brain imaging—whatever it takes. And 
also, we do preventive treatments on top of the  

01:23:56.640 --> 01:24:01.600
lifestyle [changes], that one of the problems we 
have today is a lot of people know generally, what  

01:24:01.600 --> 01:24:07.840
are good lifestyle factors. Although, I go through 
a lot more than people generally acknowledge.

01:24:07.840 --> 01:24:13.840
But they don't incorporate them because they don't 
know that they're at risk and they could change  

01:24:13.840 --> 01:24:20.160
their ... extend their health span and prevent 
that disease. So what I at least put out there,  

01:24:20.160 --> 01:24:26.960
a blueprint, is how we can use AI, because it's 
multimodal AI, with all these layers of data,  

01:24:26.960 --> 01:24:34.160
and then temporally, it's like today you 
could say if you have two protein tests,  

01:24:34.160 --> 01:24:38.635
not only are you going to have Alzheimer's, 
but within a two-year time frame when ...

01:24:38.635 --> 01:24:39.029
LEE: Yep.

01:24:39.029 --> 01:24:45.440
TOPOL: ... and if you don't change things, if 
we don't gear up … you know, we can ... we can  

01:24:45.440 --> 01:24:52.400
completely prevent this, so … or at least defer it 
for a decade or more. So that's why I'm excited,  

01:24:52.400 --> 01:24:59.120
is that we made these strides in the science of 
aging. But we haven't acknowledged the part that  

01:24:59.120 --> 01:25:07.115
doesn't require reversing aging. There's this much 
less flashy, attainable, less risky approach ...

01:25:07.115 --> 01:25:07.840
LEE: Yeah.

01:25:07.840 --> 01:25:13.760
TOPOL: ... than the one that … when you 
reverse aging, you're playing with the  

01:25:13.760 --> 01:25:18.280
hallmarks of cancer. They are like, if 
you look at the hallmarks of cancer …

01:25:18.280 --> 01:25:19.680
LEE: That has been one of the primary challenges.

01:25:19.680 --> 01:25:21.102
TOPOL: They're lined up.

01:25:21.102 --> 01:25:21.113
LEE: Yeah.

01:25:21.113 --> 01:25:24.320
TOPOL: They’re all the same, you 
know, whether it's telomeres,  

01:25:24.320 --> 01:25:30.480
or whether it's ... you know ... so this 
is the problem. I actually say in the book,  

01:25:30.480 --> 01:25:33.760
I do think one of these—we have 
so many shots on goal—one of  

01:25:33.760 --> 01:25:40.480
these reverse aging things will likely 
happen someday. But we're nowhere close.

01:25:40.480 --> 01:25:45.360
On the other hand, let's gear up. Let's do 
what we can do. Because we have these new  

01:25:45.360 --> 01:25:51.840
metrics that's ... people don't … like, when I 
read the organ clock paper from Tony Wyss-Coray  

01:25:51.840 --> 01:25:57.680
from Stanford. It was published end of ’23; 
it was the cover of Nature. It blew me away.

01:25:57.680 --> 01:25:58.100
LEE: Yeah.

01:25:58.100 --> 01:26:00.880
TOPOL: And I wrote a Substack 
[article] on it. And Tony said,  

01:26:00.880 --> 01:26:06.440
“Well, that's so nice of you.” I said, “So nice? 
This is revolutionary, you know.” [LAUGHTER] So …

01:26:06.440 --> 01:26:10.880
LEE: By the way, what's so 
interesting is, how these things,  

01:26:10.880 --> 01:26:13.760
this kind of understanding 
and AI, are coming together.

01:26:13.760 --> 01:26:14.400
TOPOL: Yes.

01:26:14.400 --> 01:26:17.480
LEE: It's almost eerie the timing of these things.

01:26:17.480 --> 01:26:21.600
TOPOL: Absolutely. Because you 
couldn't take all these layers of data,  

01:26:21.600 --> 01:26:23.742
just like we were talking about data hoarding.

01:26:23.742 --> 01:26:23.753
LEE: Yep.

01:26:23.753 --> 01:26:28.720
TOPOL: Now we have data hoarding 
on individual with no way to be  

01:26:28.720 --> 01:26:34.880
able to make these assessments 
of what level of risk, when,  

01:26:34.880 --> 01:26:41.520
what are we going to do in this individual 
to prevent that? We can do that now.

01:26:41.520 --> 01:26:47.040
We can do it today. And we could keep 
building on that. So I'm really excited  

01:26:47.040 --> 01:26:51.600
about it. I think that, you know, when 
I wrote the last book on deep medicine,  

01:26:51.600 --> 01:26:57.920
it was our overarching goal should be to 
bring back the patient-doctor relationship.  

01:26:57.920 --> 01:27:01.040
I'm an old dog, and I know what it used 
to be when I got out of medical school.

01:27:01.040 --> 01:27:09.920
It's totally ... you couldn't imagine how much 
erosion from the ’70s, ’80s to now. But now I have  

01:27:09.920 --> 01:27:15.680
a new overarching goal. I'm thinking that that 
still is really important—humanity in medicine—but  

01:27:15.680 --> 01:27:21.840
let's prevent these three ... big three diseases 
because it's an opportunity that we're not … you  

01:27:21.840 --> 01:27:28.880
know, in medicine, all my life we've been hearing 
and talking about we need to prevent diseases.

01:27:28.880 --> 01:27:36.000
Curing is much harder than prevention. And the 
economics. Oh my gosh. But we haven't done it.

01:27:36.000 --> 01:27:36.640
LEE: Yeah.

01:27:36.640 --> 01:27:41.982
TOPOL: Now we can do it. Primary prevention. 
We’d do really well. Somebody’s had heart attack.

01:27:41.982 --> 01:27:41.993
LEE: Yeah.

01:27:41.993 --> 01:27:46.080
TOPOL: Oh, we're going to get all over it. Why 
did they have a heart attack in the first place?

01:27:46.080 --> 01:27:48.480
LEE: Well, the thing that makes so 
much sense in what you're saying  

01:27:48.480 --> 01:27:53.360
is that we understand we have an 
understanding both economically and  

01:27:53.360 --> 01:27:59.760
medically that prevention is a good thing. 
And extending the concept of prevention to  

01:27:59.760 --> 01:28:03.200
these age-related conditions, I think, 
makes all the sense in the world.

01:28:03.200 --> 01:28:11.440
You know, Eric, maybe on that optimistic note, 
it’s time to wrap up this conversation. Really  

01:28:11.440 --> 01:28:18.640
appreciate you coming. Let me just brag 
in closing that I'm now the proud owner  

01:28:18.640 --> 01:28:24.600
of an autographed copy of your latest 
book, and, really, thank you for that.

01:28:24.600 --> 01:28:26.320
TOPOL: Oh, thank you. I could spend the rest of  

01:28:26.320 --> 01:28:31.040
the day talking to you. I've 
really enjoyed it. Thanks.

01:28:31.040 --> 01:28:37.120
[TRANSITION MUSIC]

01:28:37.120 --> 01:28:38.720
LEE:  

01:28:38.720 --> 01:28:42.960
For me, the biggest takeaway from 
our conversation was Eric’s supremely  

01:28:42.960 --> 01:28:49.200
optimistic predictions about what AI will 
allow us to do in much less than 10 years.

01:28:49.200 --> 01:28:55.760
You know, for me personally, I started off several 
years ago with the typical techie naivete that if  

01:28:55.760 --> 01:29:01.920
we could solve protein folding using machine 
learning, we would solve human biology. But  

01:29:01.920 --> 01:29:07.520
as I’ve gotten smarter, I’ve realized that 
things are way, way more complicated than that,  

01:29:07.520 --> 01:29:15.200
and so hearing Eric’s techno-optimism on this 
is really both heartening and so interesting.

01:29:15.200 --> 01:29:19.520
Another thing that really caught my 
attention are Eric’s views on AI in  

01:29:19.520 --> 01:29:23.760
medical diagnosis. That really 
stood out to me because within  

01:29:23.760 --> 01:29:28.240
our labs here at Microsoft Research, we 
have been doing a lot of work on this,  

01:29:28.240 --> 01:29:33.120
for example in creating foundation 
models for whole-slide digital pathology.

01:29:33.120 --> 01:29:38.480
The bottom line, though, is that biomedical 
research and development is really changing  

01:29:38.480 --> 01:29:44.320
and changing quickly. It's something that we 
thought about and wrote briefly about in our book,  

01:29:44.320 --> 01:29:48.480
but just hearing it from these three 
people gives me reason to believe that  

01:29:48.480 --> 01:29:53.920
this is going to create tremendous benefits 
in the diagnosis and treatment of disease.

01:29:54.800 --> 01:30:00.880
And in fact, I wonder now how regulators, 
such as the Food and Drug Administration  

01:30:00.880 --> 01:30:06.720
here in the United States, will be able to keep 
up with what might become a really big increase  

01:30:06.720 --> 01:30:12.000
in the number of animal and human studies 
that need to be approved. On this point,  

01:30:12.000 --> 01:30:18.160
it's clear that the FDA and other regulators will 
need to use AI to help process the likely rise in  

01:30:18.160 --> 01:30:30.863
the pace of discovery and experimentation. And 
so stay tuned for more information about that.

01:30:30.863 --> 01:30:30.880
[THEME MUSIC] 

01:30:30.880 --> 01:30:36.240
I'd like to thank Daphne, Noubar, and Eric again 
for their time and insights. And to our listeners,  

01:30:36.240 --> 01:30:40.000
thank you for joining us. There are 
several episodes left in the series,  

01:30:40.000 --> 01:30:45.200
including discussions on medical students’ 
experiences with AI and AI’s influence on  

01:30:45.200 --> 01:30:50.640
the operation of health systems and public health 
departments. We hope you'll continue to tune in.

01:30:50.640 --> 01:31:07.600
Until next time.

01:31:07.600 --> 01:31:08.694
[MUSIC FADES]