Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:00
Unexplainable is a science show about
0:03
everything we don't know. Like,
0:05
we don't know how bikes work. Get
0:07
out! Come on! We don't know where
0:10
the moon came from. Holy cow.
0:12
You've touched the moon. This is incredible.
0:14
We don't even know what life is.
0:17
No one has been able to define life, and
0:19
some people will tell you it's not possible to.
0:22
Unexplainable takes you right up to the edge of
0:24
what we know and keeps going.
0:27
New episodes every Wednesday wherever you get
0:29
your podcasts. Hello
0:33
and welcome to Raising Health, where we explore
0:35
the real challenges and enormous opportunities facing entrepreneurs
0:37
who are building the future of health. I'm
0:41
Olivia. And
0:46
I'm Chris. Today's episode is with
0:49
Kim Branson, Senior Vice President of Artificial
0:51
Intelligence and Machine Learning at GSK. He
0:53
is joined by Vijay Pandey, founding General Partner
0:56
of A16Z Bio and Health. Together, they talk
0:58
about how AI has improved drug discovery and
1:00
development, as Kim walks through all the ways
1:02
AI can be deployed in the lab. AI
1:05
has become so useful because
1:07
we've got so much more measurement technology, right?
1:09
Cheap measurement technology, so we can do generalized
1:11
sequencing cheaply, we can do RNA sequencing, we
1:13
can do single-cell type stuff. You've got this
1:15
data, you've got more measurements now, but
1:17
you can't do anything with it unless you've got
1:19
machine learning. You know, my central hypothesis has always
1:22
been medicine has advanced, this new measurement technology has
1:24
come along. And as the measurement technologies get cheaper,
1:26
we learn more. They also get into
1:28
detail about what an ideal partner looks like
1:30
and what kind of solution is attractive to
1:32
a big biopharma, making this a particularly relevant
1:35
episode for builders in biotech AI. Because it's
1:37
honestly, you're selling the solution at the data styles and people
1:39
don't care whether it's like, you know, the mechanical
1:41
Turk type thing or whatever. You
1:44
can sell a random forest if you're trying to really call
1:46
random forest. Doesn't matter, it's not a neural network. You're
1:48
listening to Raising Health from A16Z Bio and
1:50
Health. Today
1:54
we have Kim Branson, the SVP and
1:57
global head of AI and machine learning
1:59
for GSK. So Kim, thank you so much
2:01
for being here. Welcome to Raising Health. Thanks, Vijay. It's great
2:03
to be here. Maybe one place to kick it off is
2:05
that I think people would love to get to know you. And
2:08
what got you interested in the space in
2:10
general? My guess is as a kid, you
2:12
like computers, but you must have also liked
2:14
medicine and biology. How does this all start?
2:16
I actually never done any biology. Never even
2:18
thought about it until I actually got to
2:20
university. And then I discovered molecular
2:23
biology and then bacterial pathogenesis. And
2:25
you have this interplay between the immune system and bacterial
2:27
stuff. And I was sort of hooked. And
2:29
yeah, you're right. Obviously, as a kid, I had played
2:31
with computers and liked computers. But I was far more
2:33
like a math, physics kind of kid. And
2:36
once I discovered this whole world, I
2:38
was sort of hooked. And then I discovered it wasn't just
2:40
sort of blobs, blobology. There was like structure. So
2:43
then x-ray crystallography and things. And once you sort of put all
2:45
those things together, it's
2:47
sort of a very seductive thing to come into
2:49
how things really work. So I
2:52
was sort of always drawn
2:54
by the allure structure of
2:56
biology and love all those sorts of
2:59
things. And then I realized between those
3:01
long hours during incubations and things
3:03
like that, I could just go and do
3:05
this instead. And just being
3:07
computations. Yeah, doing computation and doing small molecule drug
3:09
design. So when I had sort of finished my
3:11
PhD, even during my PhD, people were like, well,
3:13
computational drug design, does that work? And the answer
3:16
was, well, yeah. And it definitely is dying to
3:18
work more. And so the people I did my
3:20
research. But you're talking about, well, that's 20 years
3:22
ago. Yeah, so we're talking like, yeah, it makes
3:25
me feel old. This is like 99. So
3:29
I was working with Joseph
3:31
Agis and Peter Coleman and Brian Smith. And
3:33
these are the people behind the first Neuromunidase
3:36
drug, which actually was designed actually with computation.
3:38
So they used a program called Goodford's Grid
3:40
to map the protein surface and work out
3:42
to put this functional group there. And
3:45
it was all based off the original sugar
3:47
molecule. But it was actually designed using
3:49
a computational method. They were believers in it
3:51
because they had built something. It became
3:53
marketed as Relenza, I think, by GSK, ironically.
3:57
So yeah. Full circle. Yeah. Actually
4:00
after that, you went on to go to some
4:02
startups. Yeah. And also go on to pharma. Curious
4:04
to get your take on one versus other, how
4:07
you see that ecosystem. I did some sense of
4:09
time at Vertex. And I wrote a Vertex in
4:12
biology as you know. The best model we can build has
4:14
a figure of merit of accuracy, like 0.6, 0.7. If
4:17
it's 0.9, it looks great. There's
4:20
something wrong. This is pre-bing.
4:22
This is the ideas of Power Set and Bleco, if
4:24
you remember that, and a whole bunch of search things.
4:26
So I went. Some friends of mine were building a
4:28
search engine. And so I went to join
4:30
them because I wanted to play with large data
4:33
sets. And I learned a lot about
4:35
doing machine learning and things on very large data sets. And
4:39
that was really exciting because
4:41
the apps are up. There's not many encumbrances you
4:43
just get on with things. But then
4:45
you're inhibited by scale and capital and everything else
4:47
and those other types of things. Whereas
4:50
at a large company, there
4:52
is a lot more capital. But
4:54
also, it's a very big ecosystem. A lot of people are doing
4:56
various things. The trick is how you direct it. They
5:00
have different strengths and weaknesses. And for
5:02
me, I guess I could
5:05
bring what I learned from one to the other and
5:08
then back again. But once I
5:10
started the startup world, I stayed
5:12
in that bet for three companies.
5:14
For me, I still really
5:17
like that feeling of pace and
5:19
urgency. Yeah, in the startup. Yeah, yeah.
5:21
And if you carry that through, I
5:24
think in a large company, it's very easy to sit
5:26
back and go with ebb and flow. You
5:28
can do that if you do it today. If
5:31
someone else delays three days, then it doesn't delay a
5:33
month at the end. So do it,
5:35
do it now kind of thing, energy of startups.
5:37
And you recruit different type of people and things
5:39
like that. So what drew you to
5:41
GSK? What brought you there? So GSK was
5:43
interesting. So I was at Genentech at the
5:45
time. And I had a lot of friends at
5:47
GSK. And they
5:49
were wanting to Do AI and things. So They
5:51
asked me, how should we organize this? Where should we put someone?
5:54
They were dead set on a particular location. And at the time,
5:56
I was like, you're not going to find the people you want
5:58
there. And you should organize it this way. And
6:00
I was quite resisted going to Gsk because
6:02
if you know outside perceptions and you see
6:05
the same people at conferences that last twenty
6:07
years complaining about the same thing I want
6:09
my class right to be company for Yeah
6:11
yeah and so he eventually this is like
6:13
wages com an interview and so I went
6:15
to interview and I said have met how
6:17
I met some. People. I realize
6:19
actually. It's. Kind of very
6:22
different rides and then my and like realize that
6:24
they really a waltz in a phase of this
6:26
telling us up inside out these become don't reinvent
6:28
itself very often your hard to do yeah yeah
6:30
but they were really do and my were really
6:32
serious about doing it and so you know it's
6:34
an amazing how and I knew. He
6:37
how of the A through Daphne and other people
6:39
and you knew he was with and he gets
6:41
to write some python Ryan we get we get
6:43
his identity of oh yeah yeah he going to
6:45
see as an accident nearby Bananas well so he
6:47
can really right right coats and really like this
6:49
guy that she went missing people He knows what
6:52
it takes to do things and they're fully investing
6:54
in the things I thought were very important the
6:56
time which was you know these last year databases
6:58
coming on mine and vintage Nymex or he was
7:00
what way ahead in the in some of the
7:02
to give up crisper unlike the gene pool of
7:05
base generate data and saying we can discuss exclusion
7:07
data you can see the trend. So began the
7:09
machinery the make sentimental and so they weren't sort
7:11
of just like dabbling in like will have one
7:13
group and will do some contracts like move to
7:15
establish entire thing and it'll be caught. Why strategy
7:17
so how will that seems like an opportunity and
7:20
I you know I guess I was thinking that
7:22
time of doing another company or something else I
7:24
the worst case scenario did that on the job
7:26
process like it's not very hot and in an
7:28
interesting way there was a book running you probably
7:30
know that he were running up a book that
7:33
how many how long I would lost their like
7:35
to our really mans like two years interests. I
7:37
beat the odds. Now it's five years this
7:39
Monday. Know that that I've been there. You
7:41
go short squeeze on a if I if
7:44
I if an hour nice as I. Say
7:48
I've stayed on any. we built a lot. I
7:50
think of the I know How to Do things
7:52
in smaller companies. He had a build thing we can
7:54
track the peep when everything else but the next release
7:56
of challenges. How do you take a large organization create
7:59
something within a bit. They'd bring everybody along
8:01
for the ride and things like that, which
8:03
is a new challenge it was. And how
8:05
do you do that? I mean that's a
8:07
great topic I guess and Outlaw isn't just
8:09
for seep uses the company so that ratio
8:11
flight computation to communication right is really important.
8:14
So what I realize is that if. You
8:16
have to spend it's a long time. Explained
8:18
my what he seems intuitively obvious to you
8:20
to a large organization. has he been a
8:23
market A very busy people they will but
8:25
jobs to do and are coming to different
8:27
backgrounds to make some. and like most drug
8:29
discovery takes people from. All. Sammy
8:31
different disciplines together right path manufacture muthee like
8:33
it's entire thing. It's is complex endeavors anything
8:35
else and so they don't have had the
8:37
same background and training and and also different
8:39
ages. rotten never like tech only change a
8:41
lot So you got people living in different
8:44
stages that of what their experience with and
8:46
some I'm skeptical. We've got some in a
8:48
worked as as we work and compare it
8:50
to think so you've got people that on.
8:52
It. For a classic innovate and Emma when was not
8:54
very good. Now I'm like yeah but I say
8:56
nothing of the right a change. Yeah, I forgot
8:58
Moore's law, yeah you only got then you got
9:01
true believers near in of poverty those thing so.
9:03
What? It is about can expensing
9:05
people doing lot of messaging but
9:08
descend point. On I discovered you
9:10
can talk a lot. When. It
9:12
easy to south and build for bit so he
9:14
does basically like Caymans had said no we're not
9:16
working on this of the things cradle to show
9:19
for bubble built in the organization boat a whole
9:21
bunch of capability and then you at this phase
9:23
now it sort of the installation phase of the
9:25
technology right? yep where you now going to plug
9:27
into We've done with a pilot from things in
9:30
a ones that actually make it part of the
9:32
process but I think that's where you have to
9:34
communicate to communicate and communicate right and it takes
9:36
a long time cause is also geographically distributed time
9:38
zones things like that so he be surprise of
9:41
mouth time it takes. For a messy sir to
9:43
percolate but eventually if one's ever gets behind it
9:45
and you know then you give tenure should take
9:47
and then it on stuff. a dog fight for
9:49
the thought of capital One data and well that's
9:51
when when I have a farmer you've got all
9:54
that I have any discipline you want. You can
9:56
find an expert now and if I if they
9:58
don't know the answer they know decks that. who
10:00
does. I think the question now is like okay
10:02
we see around us but how is it going
10:04
to affect drug design I think was really cool
10:07
by what you're doing is you're doing this today
10:09
and was you could give some examples. Yeah so
10:11
I think everybody thinks or with I are successful.
10:13
You entering the disease and it will say this
10:15
talk and pop out the molecule. Make it and
10:18
we. And noom you know we went.
10:20
Even each test will test inventions humans right I
10:22
always bromides nice. I hear that last maybe don't
10:24
have to the me best years are well they
10:26
are so I think I think what we're seeing
10:28
is like. Your our group books or
10:30
the way across the process right? We have an
10:32
enterprise focus on. The. First thing is like
10:34
want to make the medicine against. right? You
10:37
could design the best dragon well with the
10:39
best pharmacokinetics and safety profile of the else
10:41
but if you're much lane among thing is
10:43
that the have a clinical effect my suffer
10:45
thing is it's the time we have these
10:47
last night databases of people sequenced and the
10:49
medical records and things like that and we
10:51
look for here people the disease here people
10:53
don't have the seasons are what's different about
10:55
them may lots of differences but what difference
10:57
is a really significant and in are they
10:59
drive ins disease so you might festival do
11:01
people that are diagnosed and not diagnosed net
11:03
one way of doing things. Or. We
11:05
can. Use. Machine. Then again,
11:07
so he isn't she. Many get on clinically meeting
11:09
say a bunch of people that may have a
11:11
different range of a site say live in management
11:13
all scarring, a liver or kidney function and rather
11:15
having a person score that by hand. we can
11:17
have a I model gone to that. The
11:21
gives us what to call a continuous train Him
11:23
it can take that information and do d was
11:25
right to so by you'd wake up was De
11:27
Vito. Why does Ossetia exactly right now we have
11:29
some other variants on that that were were interested
11:31
as well. My. But once.
11:33
You have that he didn't need were one of
11:35
the genetic variant do like biologically and the first
11:37
question is what cell types as acting that it
11:39
could be like several genes away and right so
11:41
you know predict which is which were the acting
11:43
and then it is a cave. what is it
11:46
doing. Is it making more of
11:48
the messenger or less the messenger or changing
11:50
spicing, right? So you get a different form
11:52
of the party And so we have methods
11:54
that they're missing money. methods that will predict
11:56
the directs. Now they are and so that's
11:58
in use in production to. The
12:00
health isn't it for the variants I and
12:03
so that helps as I do York I
12:05
dissing make movies Protein Pp was into had
12:07
this disease. Therefore, This gives a
12:09
clue on how much like the to these
12:11
right We have a. Lot
12:14
of different methods for looking at sell it,
12:16
imaging, sitting if you have occurred and things
12:18
like that we been predicting me. effect of
12:20
a small molecule can't like it on a
12:22
cellular phenotype. Arm. We have an
12:24
active learning systems for discovery and doing experiments
12:26
in biological models of disease and Sir Robin
12:29
making a small molecule tumult to molecule. Now
12:31
we can actually to sputtered that we just
12:33
turn up and down that the regime know
12:35
I any do paralyzes lung anymore. the one
12:37
and Bolland krispies him go to Talon. We
12:39
can turn up and down the things as
12:41
well when a continuous basis. We can go
12:44
from like small amounts, a large amount of
12:46
protein or tenant, turn it down slightly and
12:48
see what happens. He can modulate things like
12:50
that. And so we have a whole active
12:52
money system where we. Feed him he's are
12:54
they from here are hypothesis objects He's a
12:56
stuff on which to his like just trust
12:58
me I know happen as fast as you
13:00
get that in biology can't have you experienced
13:02
any. We can't do a genome wide screen
13:04
just turn it to try everything. So we
13:06
use this model that says okay I wanna
13:08
say i want my my. My
13:11
aren't expression profile to look like there's some
13:13
like some proteomic thing or maybe a functional
13:15
as I look like this and ultimately not
13:18
toxic and a tiger that's tackle like a
13:20
my Kamala against right So you michael those
13:22
things in. And. It
13:24
will do around of experimentation say well he has
13:26
information back. And in
13:28
by some of the I might saying by some
13:31
I conduct connectivity understanding. On. My second
13:33
a walk up there me take this guy
13:35
cause the common ancestry two parties are and
13:37
that is of the signal transduction works that's
13:39
about twenty percent faster we shine and doing
13:41
a random screen and as a going to
13:43
the clinic we use the seen running on
13:45
we accomplish pathology so we be writing competition
13:47
pathology for while we have a even as
13:49
he has a professorship of companies with alleging
13:51
that Geico to see Razvan hum who works
13:54
as closely and we use that the. First.
13:56
Of all we know, we do a lot of things with people's I you
13:58
know is that I've expressed in this issue. Going on holiday
14:00
right? And that's just yeah. you're staying with an
14:02
antibody and people are humor pathologist or Clinton basically
14:05
the I the about sixty three percent of the
14:07
slightest this you give it to few different people
14:09
inside as expressed high medium or like more when
14:11
the see money I can say how much it
14:13
express with counting brown pixels it's pretty easy Mac
14:15
my pillow thought can say would sell type it
14:18
is. right? And so then you
14:20
can the i just met office's gate. Frida.
14:22
Point is hop to do other things but then
14:24
once you've got that us a concise to get
14:26
responses me trawling and they will now which cell
14:29
types expressed in to respond more unique you got
14:31
more day in and use it right on the
14:33
model so there's things that that and we have.
14:35
We did a very heavily instrument advice you trolls
14:37
we measured lot more. Part L makes things and
14:39
everything else and before and very expensive ailments and
14:41
like will have you noticed. Me useful might. Not
14:44
most. Of it won't but some of will
14:47
be and then it would And we were would
14:49
look at that and analyze that data and find
14:51
this some sort of people that we know if
14:53
we low their vows as as entrances, level by
14:55
up as a function. Cute right and we know
14:57
which group it is. Inigo Okay now I can
14:59
find him as the brings in that state in
15:01
a country and the destroy them. My good to
15:03
go also for the audience. could you give a
15:05
sense for the delta between like which you can
15:07
do with Ml now a Gsk vs maybe the
15:09
way this would have been done ten years ago.
15:11
And. The Delta in terms of time or dollars
15:13
a pit wherever? Yes, yeah, I mean certainly for
15:16
some of these things you would have to do.
15:19
A lot more experiments, right?
15:21
So the genetics. Privee.
15:24
Didn't know which way the bearing was. Well, you'd have
15:26
to go an. Engineer a system and
15:28
make a model system and ten more been
15:30
up and had more went down all over.
15:32
express it or not and then make whole
15:34
bunch of Sergeant Recounts arm and all the
15:36
things about making continuous traits for discovering. Wanna
15:38
see a whole bunch? A radiologist has scored
15:40
a whole bunch of things manually and did
15:42
it all right. Net would take a long
15:44
time for that's a very easy paolo cost
15:46
right? Eat We can calculate the complex you
15:48
some of these are the things is. You
15:51
can go through. Fum. A very
15:54
small and heidi space. and try and find
15:56
out some Carl isn't that something like that that?
15:58
He just gives them an advantage. No,
16:00
I need a New York. To do
16:02
it reliably making some things a very complex. Ryan
16:04
said a happy example is in Iowa and. It's
16:07
a base in work in like an Orang how
16:09
to type thing but you doing a lot of
16:12
it's compression type things ryan said it is thing
16:14
because. I. I become so useful.
16:17
Because. He got some more measurement technology
16:19
right seat measurement technology so he can
16:21
do. Beginning. July sequencing city we can
16:23
do our own i seek with using sulfide
16:25
stuff you can't be state. He got more
16:27
measurements now. But. You can't do the
16:29
with unless you got machine learning and I'm
16:32
in my central hypothesis always been medicines as
16:34
arts as new Measures attack on his come
16:36
along along and as amendment technologies get cheaper.
16:39
I'm we learn more so if you think about
16:41
cardiology right yes once cat cats that are measuring
16:43
things in the heart of a people can see
16:45
that. My story the first consider his bunkers he
16:47
does himself by oh yeah yeah yeah it is
16:49
like our that it would myself not as outs
16:51
absolutely kind of old school stuff is identify the
16:53
i'd of a have area are able to hum
16:55
yeah the guy the yeah I mean about the
16:57
pylori got a idea stranger my god what am
17:00
I think I might as an eye and they
17:02
said it's high they had a hypothesis back to
17:04
of they were actually so I know who they
17:06
want to try business and were no theoretical perfectly
17:08
or it outlasts now I'm but. Yeah I think
17:10
they're gonna go far as the I believe he
17:12
did their deaths as were enough experimentation and try
17:14
this at home Now I'm but from that either
17:16
you what blood pressure cos you're things now we've
17:18
got these right, that's enough pulse. My hunting for
17:21
that as I as I got cheaper we learn
17:23
more about cardiology right? And then that led to
17:25
discovery. So we've got more has been technology. Based.
17:27
In such high dimensionality on people that like the
17:30
item makes sense of it's an array of like
17:32
whole bunch of expression changes which suspects were around.
17:34
Baseline bus is a disease patients yeah right like
17:36
who can make sense of that So this we
17:39
could deny that we would if without machinery when
17:41
we would know what to do with it. Now
17:43
now I we have the various a linear things
17:45
and take up the most obvious stuff which might
17:47
be wrong. It's hard to compare in the event
17:50
of the data. Respect nine the change things Yeah
17:52
yeah I'm curious. Love our listeners are some people
17:54
that are doing start from the start businesses and
17:56
would have you learned. Would you take to start up
17:59
If you were to start. Now. Did. You
18:01
figure like a I sat In general, I'm. It's.
18:03
About data. It's an end
18:05
in itself about i'm unique data so what you want
18:08
is I think of a com prob my how I
18:10
see is people try and build something as like they
18:12
don't having that he's I'm on a set up to
18:14
do and I for we don't have exactly the data
18:16
we've got. This sort of said you saw this is
18:18
probably good as the people forget that the data they've
18:20
got is sometimes a proxy for the real thing and
18:23
what you say that's not the did that they doesn't
18:25
have the aren't easy going. Generate the data and actually
18:27
be under generate the data is gonna be critical cause
18:29
you're gonna run into thing at some point you run
18:31
out of and you need to generate the data to
18:33
build up and a bottle and and and. That also
18:35
becomes a competitive mode. Have some reckless you'd
18:37
have. it's not that you can generate. Plus
18:40
the public staff they've got in are you
18:42
bringing twice days that hate? ideally someone pc
18:44
the jury that they get a okay see
18:46
the data and in pays you to buy
18:48
them auto battle system here that that just
18:50
won't do that they would you just to
18:52
be clear, haven't But yeah, that's that's because
18:54
the data aspect is key in the joint
18:56
bid. Did it get more data right? Like
18:58
invest in that cause honestly more data. With.
19:00
A simple I were them is an easy sell, easy
19:03
install and like data wouldn't die in a look at
19:05
the Netflix sounds how was that One million different of
19:07
I Don't Daves to the table not with a smart
19:09
our them right us So I think that's a key
19:11
lesson is immense in there and then and then think
19:13
about the Ml and the fancy i wear them and
19:16
things that that's of wow look at lot of Com
19:18
is a come to asa look at like what unique
19:20
dated you have met and what's your been a degenerate
19:22
and what scaled you have I that I don't for
19:24
example I mean if you're pitches dislike all we've got
19:26
smart people give us your that it would be cool
19:29
stuff on my blog of smart people to rise. Of
19:31
the happened what was in what difference is
19:33
the data in your mind. I.
19:35
Think is the. Either. You've
19:38
got a. It's uniquely
19:40
suited to understand so it's clean, right?
19:42
You got this batch effects. You can
19:44
generate more that and you can control
19:46
the sampling aspects of it. I think
19:48
those things like a key you. Could.
19:50
Then basically understand how your method
19:52
might behave under different scenarios where
19:55
what would which may have rocket
19:57
us and biases and as systems
19:59
example interesting. You're in the abstract, those are
20:01
kind of the key things you would wanna kind of
20:03
point and then we you got that in the Ml
20:05
thing doesn't really matter because it's almost. you sign the
20:07
solution in the days of and people care whether it's
20:09
like you know the the mechanical turk type thing or
20:12
whatever we as you can sell a random forest minority
20:14
cool random forest if you know it doesn't as on
20:16
a new network or whatever L hundred and one of
20:18
our assignments I goes along rak get their hands if
20:20
they did all on the horizon for the place you
20:22
visit based on average to start with every night rider
20:25
yes a she always compare your information gain of of
20:27
a random forest that could be depressing from a. Skill
20:30
or either that or or in any of the
20:32
other simple method suggests even desserts a similar neural
20:34
net. It was logistic regression basically. yeah yeah yeah
20:37
I mean that same thing I've it's it's funny
20:39
actually it's how I think that will go full
20:41
circle. By the way there's of his love of
20:43
the a complex stuff on we'll see what happens.
20:46
Yeah I think we people realize that I'm on.
20:48
This is great paper like by David Hand by
20:50
which we may have said long time ago by
20:52
the could classify a complex the an illusion of
20:54
progress. It's very old square malware he looks at
20:57
some you see of on stuff and that end
20:59
of a sophisticated. Mythical Macinnis Vm but it's
21:01
like this year disposition any discomfort analysis at
21:03
the are a bunch of things and he
21:05
very toy datasets but saying that like facing
21:07
the simplest of gets you there and that
21:09
extra five percent gains you need with extra
21:11
complexity. I it's and still true now I
21:13
think they'll think the thought his work out
21:15
how good the thing needs to be before
21:18
you start. So Tokyo Macula got like if
21:20
he could do this, do these things yes
21:22
I would use it and then go work
21:24
out to do that don't build a thing
21:26
and and sidewalk got that with an hour
21:28
to work it out and eat. Maybe you're
21:30
far off or maybe there are some criteria
21:32
right? and I were you. It's top secret,
21:34
the pure. you know what's stopping when to
21:36
start shipping and but so to start had
21:38
to differentiate data. yep at what would you
21:40
wanna see in the algorithm to get excited
21:43
I want to see. Well
21:45
characterized robustness and reliability criteria. Right arm
21:47
don't give me to point estimates you
21:49
me a point estimate with some measure
21:52
of confidence as her right like you
21:54
gotta have that are many minutes se
21:56
tricky to do sometimes they are higher
21:58
than I can still. I mean
22:00
like I still. Think
22:02
that's what I'm missing. My papers out there
22:04
is like well my things better he is
22:06
my the his my table his my stuff
22:08
in ball which is naturally by the my
22:10
stuff and by the way different of the
22:12
third different place for my criticism by this
22:14
other thing is like the lot of people
22:16
stuff the ten percent difference that like I
22:19
don't care about worth the cost me by
22:21
inspired know is if you like point one
22:23
in asia not gonna not so yeah you
22:25
want the I'm the precision cove right I'm
22:27
is really mirza of a you say but either
22:29
way out what you need to be at
22:31
home to make it many be different and then
22:33
how will the engineered and in time plug
22:35
into my system housing work right out the in
22:37
the installation stuff could often knit the people
22:39
who use it in a company in in
22:41
this be will have to operate and run and
22:44
had remote control for model drift in everything
22:46
else and so you're. Selling.
22:48
To a user. But the people
22:50
buying and installing it that are also part the
22:52
decision process are like the engineering thoughts ego thing
22:54
to a cafe, how we make it easy for
22:56
them to use and I would just they can
22:58
be around people like stick in the Cause to
23:00
get my cloud. You. Know
23:02
that we have run on PRAM like if the
23:04
think of different ways to bring it to the
23:06
market and that depends on the industry. Like
23:09
the more ways you habit the make it easy
23:11
if he would integrate and monitoring watson it's security
23:13
than let's stop the is it is for them
23:15
for to be fixed with thing. I.
23:17
Know you still my coding same time so
23:19
I know you like playing with a eyes
23:21
or and what have you done your own
23:23
hands Recently moved and the one thing I
23:25
bear with me is a i can these
23:27
emails reports fine I'm over the various types
23:29
of things and like eighty one question when
23:32
I was one thing it's like I come
23:34
by their at all so. I
23:36
got realize like I just thing that like
23:38
dumpster might these emails clear whether since then
23:40
title onto a photo my laptop and I'm
23:42
on the Lm as my quest honor and
23:44
in it for me via email to sense
23:47
to me write an essay like that's one
23:49
thing I built recently. I'm. In. I'm
23:51
pretty interested in Ny and of the long
23:53
contacts windows the the stop us as like
23:55
I'm. Do. Need a long contacts minute Fatah
23:57
spanning Or do you need a whole bunch of specialist June.
24:00
The lens we're in an age of for dry
24:02
yeah yeah I'm going to search yes no yes
24:04
and look such as are things such as fundamentally
24:06
changed because you know what in like easy like
24:08
find me to document I'll read it now it's
24:10
like he's my question you read the document for
24:13
me and telling my answer is obvious nissl version
24:15
of actually this Lm for reasoning over multiple scientific
24:17
documents that I write myself russians impact on our
24:19
the prototype and put it together and now it's
24:21
a tie I still think is do have that
24:23
kind of build things I think also to. It.
24:26
Keeps you think emails aspects of also
24:28
think it also helps you know when
24:31
your mum bedroom for search for the
24:33
company. Ride. You're kicking the tires and
24:35
it regularly have you found it difficult to do something that
24:37
now that he would do as per be slowing them down
24:39
says nothing a lot about had I spit bit of the
24:41
duration Zoc with have. Any Mccroskey is Kate
24:44
that has eluded me running chi that he
24:46
would access data and get data and think
24:48
about things right in Iowa can take some
24:50
three weeks to get a thing to answer
24:52
simple question than yeah, that's gonna modify down
24:55
the chain, you know, and you're after another
24:57
quarter what he called us before they've done
24:59
something right. Now. Okay, so
25:01
let's it's five years from now. Yes, and we're
25:03
doing other podcasts of Will Do and sooner. But
25:05
my guess is five years. What are the examples
25:08
that you're excited about them? that you've done. A.
25:10
Guy obviously that hasn't happened yet. sometimes I
25:12
think like a month would you think really
25:14
think you'll get I think will have most
25:17
but I will cope. Computational methods every drug
25:19
be have an aunt were have software the
25:21
sits around it adding it can be true
25:23
of everybody is making we're going to see
25:25
it does suck at it will sit around
25:27
every dragon with say he's do that I'm
25:29
what their coffee to be at baseline. More.
25:32
Likely what they like to be. plus one
25:34
cycling drug because it's very difficult to predict
25:36
people's responses. A Python maybe buy them will
25:38
be better at it. I think we'll do
25:40
we will know a lot more about I'm
25:42
I I I think we're really understand how
25:44
on this image therapy Gsk Famous really action
25:46
a mean programming company right vaccine that way
25:48
the program immune system. right? And
25:51
with single of different vaccines I think we
25:53
will understand really the whole the my i
25:55
always like twenties and a patient's respond by
25:57
and is amazing response but with twenty percent
25:59
I'm not. That what I mean disease. I
26:01
think we'll see get a lot more aspect
26:03
with software but I think also the ages
26:06
single cell and perturbation genomics is just getting
26:08
started. Or I like it when
26:10
it's we didn't give case be doing excellent
26:12
Would a to growth rise if I started.
26:14
You know and like with doing what we're
26:16
doing when we are generating large scale put
26:19
a basin datasets purely to become like either
26:21
lookup tables. I don't have a dune experiment.
26:23
right? But more important now an infant stable
26:25
kind thirty minutes from and for I like the
26:27
Human Genome Project becomes of lookup table Yeah yeah
26:30
exactly right. So you building most of the things
26:32
I think in five years time what will happen.
26:34
I think we'll be doing. If
26:36
you're extremis more informed of experiments but
26:39
I think also the loops into observation
26:41
cohorts limited things and people were not
26:43
things you management which is learning about
26:45
the disease or will it be running
26:47
trials limit disease. Add to. It
26:49
public, private and a mix of data so I
26:51
think that's going to be full scheme ever been
26:53
we doing that when I think you know this
26:55
whole. Wheat thing of to these
26:58
heterogeneity will actually sought to really understand what
27:00
that means night and also kind of Iraqi
27:02
sort of serious obese decipher. But yeah I
27:04
think we'll know a little bit more about
27:06
what's happening. I'm and this is where I
27:08
think interestingly. Ml. Will start to
27:10
collide with really the. You
27:13
know that kind of mechanistic modeling abolish the we
27:15
know a lot about medicine. we know organ systems
27:17
and things like that. We have to start with
27:19
just gene expression in the thing and until our
27:21
our the nothing about by the rights to the
27:24
person the south things interact we can have these
27:26
the structure price so I'm think this is where
27:28
that sort of systems mass an approach would become
27:30
very useful as well so we will have a
27:33
lot more morals I think that be used in
27:35
place for discovery and systems now as well To
27:37
witty thing is sort of the rate limiting step
27:39
in getting nailed Advancing this in our he could
27:42
be people. To be money? well and a that me
27:44
up the try to answer it. Is data
27:46
but the real answer is day with outcomes right?
27:48
Say like I can deny all the pet of
27:50
a snow day or I want the what I
27:52
really want is like. In. A sample
27:54
from the wealthy people with humble think with off
27:56
I want as what happens when I'm I was
27:59
a clinical trial. Curbing Someone right now we're
28:01
therapy with we Can Happen and Weather conditions
28:03
Yeah, at which? Yeah. And I think that's
28:05
where we want that. we need the Anthem
28:07
dataset. Outcome Daters is is the goals
28:09
and they have Fry is the rare estate
28:11
or that's really like one divine to some
28:13
zoo. Com is hop as I can generate
28:15
data in humans at scale him safe ethical
28:18
manner. That data past if you measure was
28:20
different things. We start instantly. Charles Moore. To.
28:22
Me to complete ways to be brick dragon.
28:24
Two hundred people who eat more of them
28:26
closely which emits everything we can write like
28:28
and bring that back and there is a
28:30
patient burden. You can overload them and is
28:32
a was attending between anniversaries do that right
28:34
fashioned. but that data with outcomes that's the
28:37
right them. except for the know what we
28:39
did them you bomb as as as as
28:41
that and you know how they do these
28:43
progress on what happened to them So how
28:45
can we sort of minimise Obama? That's where
28:47
I think that we either need to create
28:49
larger public private consortium or public thing like
28:51
that. In a way everybody. Wants to do is
28:53
like out to can I have odious case clinical data
28:55
for this trial and I mean if a company like
28:57
know cause I have a psyche eyes and of a
29:00
my uni committee but I have no I'm I wanted
29:02
on using it right now to develop things but what
29:04
we actually need is I think we need to do
29:06
large scale of xvi so cohorts the palm and there
29:08
are these things exist now but the palm is what
29:10
they measure is limited by the by a had an
29:13
ipod the had the time. We. Use you
29:15
gonna look at something else and I can
29:17
walk a new thing. How I measured that
29:19
when we need to bank samples my and
29:21
sometimes you in a bank the samples and
29:23
kind of invoke slackers law like a to
29:25
become proceed from bad to analyzing and then
29:27
my that is true. I guess it's Richard.
29:29
Rider is is and sometimes you can't do
29:31
with friends you middlesex we should be establishing
29:34
like it's not just genetic sequencing, an electronic
29:36
medical records, Is. Sexy can possible, but
29:38
also we'd be falling people over time
29:40
like we're even really know what the
29:42
mean system looks like overtime. right?
29:45
It's kind of bunkers I We just don't
29:47
know what the average evolution immune system does.
29:49
That's the right limiting steps and then you've
29:51
got treatments and you could outcomes. I for
29:53
those people, If the missing
29:56
link me the or that the drug discovery mean
29:58
gentle had more those sorts of thing I I
30:00
think we would we will be in a better
30:02
place and I think that's was going to
30:04
come by but that's actually what honestly I think
30:07
is the work of you know, government and industry.
30:09
the phone, those sorts of things that maybe there's
30:11
a time exclusive access and things like that but
30:13
eventually they made me pay. If I become public
30:16
resources they become a common good a new different
30:18
ways people trying to construct those for I think
30:20
it if you look away machine They've had
30:22
message strides as when is cheap by electing data
30:24
from a suburb thirty Me that was standing right.
30:27
Yes, the good advice of the thing for this
30:29
datasets. I'm we. The company that we we do
30:31
put out they assess machine learning challenges that Europe's
30:33
and things like that we've done them. You know
30:36
the genius guy challenge her pet have a sense.
30:39
Of he does a few of them. You can see the munchies.
30:41
could I mean the put out a prize of things like that
30:43
and he could make you can make real money. So Kim thanks
30:45
so much for joining us! Reason: it's been lot of fun with
30:48
it. Think.
30:56
You been listening to raising? The
30:58
posted and produced by Me Prostitution.
31:00
And me Olivia Lab with the help of
31:03
the bio and health keen a sub Kinsey
31:05
the sell its edited by Still Hugs as
31:07
if you want to suggest topics or future
31:09
shows even reach us at Raising Health or
31:11
a sixteen V.com Finally, please rate and subscribe
31:13
to our set. The
31:17
content here is for informational purposes only
31:19
should not be taken as lethal business
31:21
tax and doesn't advice or be used
31:23
to evaluate any investment. Our security and
31:25
is not directed at any investors are
31:27
potential investors and any and sixteen the
31:29
fund. Please note that a six pm
31:31
the and it's affiliates maintain investments in
31:33
the company's discussed in this podcast. for
31:35
more details including a link to our
31:37
investments.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More