Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:01
Actually 90% of all drugs fail . So
0:03
basically you're spending five years
0:05
and then you have another five years and 90%
0:08
of the time will fail .
0:09
In a world where 90% of clinical
0:11
drug development falters , companies
0:14
invest a decade for the slim
0:16
chance of success . We
0:18
dive into this high stakes
0:21
journey with Dr Max
0:23
Jacobs , CEO
0:25
and co-founder of DeepMirror
0:27
.
0:28
Because , on one end , you can use machine
0:30
learning , artificial intelligence , to
0:32
take your past data on
0:34
drug compounds that you tested against
0:37
the target and then basically try to
0:39
predict other compounds that might even be more potent
0:41
.
0:42
AI is compressing years of research
0:45
into weeks , tackling
0:47
vast datasets and promising
0:49
to accelerate pre-clinical work
0:51
, but the challenge is steep
0:53
. Studying clinical outcomes
0:56
remains a formidable task
0:58
.
0:59
Individual drug . You have to first make it
1:01
, test it , and so on . It might
1:03
take two , three weeks or so . So testing
1:05
tens of thousands is really not possible
1:07
.
1:09
The goal , Level the
1:11
playing field , increase competition
1:13
and challenge norms . With
1:17
technologies from the 1960s still
1:19
prevalent , the industry is
1:21
ripe for disruption .
1:23
The pacemaker has remained the same from 1960
1:26
to now . Largely , it seems
1:28
like there is a certain level of bias
1:30
that sticks around
1:32
that is very hard
1:34
to think completely
1:36
out of the box .
1:38
From small molecules to advanced therapies
1:40
. We're at the cusp of an explosion
1:43
of ideas and innovation .
1:45
The biggest challenge for any software
1:47
company in this place has always
1:49
been to resist
1:52
digging for gold almost
1:54
.
1:55
Join us on Things of Change podcast
1:57
for a deep dive into the future of
1:59
drug discovery and healthcare
2:01
innovation . If
2:08
you had known how important the technology
2:10
economy was 20 years ago , would
2:13
you have done things differently ? The
2:16
internet , cell phones , the cloud
2:18
and data Things
2:21
have changed . We're here to talk about
2:23
it , hi .
2:25
I'm Jed , hi , I'm Shikhar
2:27
. Welcome to Things of Change
2:29
, your new economics and
2:31
technology podcast . Healthcare
2:40
in general is hard . Biotech
2:43
, healthcare , life sciences industries
2:45
are one of those industries where
2:47
there's just a massive amount of
2:49
data . These are far
2:51
too long Just because experimentation
2:54
. It costs a lot and you need to be precise
2:56
because ultimately , the end users
2:58
are the patients and consumers are there . Having
3:02
technology come in and improve
3:05
and influence this level
3:07
of data that researchers , scientists
3:10
and companies biotech companies , small
3:12
and large operate with is
3:15
a big boon . Your algorithms
3:17
could be like your buddy that could just
3:19
help uncover some patterns that
3:21
you were not seeing before , or
3:23
even identify potential
3:26
drug targets . It
3:28
seems like healthcare in
3:30
general is one of those industries that could massively
3:33
be influenced by the
3:35
wave of AI that we're seeing today . Today
3:38
, we're super excited to have Dr
3:41
Max Jacobs , co-founder
3:43
and CEO of DeepMirror , a health
3:45
tech company building
3:47
software , ai software platform
3:50
that helps researchers accelerate
3:53
experimentation testing
3:55
and , by doing so , unlock
3:58
creativity . Think of chat , gpt
4:01
or GitHub co-pilot , but co-pilot
4:03
for your research , like
4:05
research GPT that names
4:07
you to be trademarked .
4:08
Thank you for having me . We're excited to be here
4:10
, guys .
4:12
I touched on it a bit , but what are
4:14
your thoughts as to why this is
4:16
needed in healthcare in general
4:18
?
4:18
Yeah , I think it's very important to
4:20
maybe appreciate that in relationship
4:23
to other industries in health
4:25
tech and biotech and pharma these
4:28
data-driven approaches
4:31
are still very experimental
4:33
. Some of the larger companies
4:36
might have some internal research teams
4:38
already that do this kind of work
4:40
, basically providing algorithmic
4:43
solutions to other parts of the companies
4:45
, but it's really
4:47
not yet in the hand of end users
4:49
really so these kind
4:51
of data-driven technologies . It
4:54
is really at a stage now where people maybe
4:56
until a few years ago didn't really believe
4:59
that there could be some sort of vertical
5:01
solution in the industry , because everything is , as you
5:03
said , so complex , so difficult and
5:05
so specialized to the particular application
5:08
that nothing can really help
5:10
everyone in a way . But
5:12
now we're seeing more and more approaches and
5:15
attempts to actually do this
5:17
, to bring these kind of technologies
5:19
that look maybe very difficult into
5:22
the hands of people . I think you mentioned chatGPT
5:25
. That has been demonstrated by especially
5:27
these kind of approaches . Prior
5:29
to that , people worked with large language models
5:31
chatGPT is based on , but
5:34
these were technical people . It
5:36
didn't really get to any sort of end
5:38
users , but this kind of leap
5:41
of design almost made
5:43
it possible for the first time . This is what we're
5:45
trying to achieve as well , like having
5:47
a little leap of design and user experience
5:49
to finally bring AI-powered
5:52
algorithms to all these labs
5:54
that are not really using them at
5:56
this stage because , de facto , it's
5:59
in the hand of a select few companies in
6:01
some way that have successfully applied
6:03
it . But apart
6:05
from that , most companies are like basically
6:07
thinking oh , maybe I should get in
6:10
on it , but I'm not sure how yet
6:12
.
6:13
Yeah , one of the things we were really curious about
6:15
and you talked about it a little bit was how does
6:17
current state look ? What are the
6:19
kind of issues that researchers are facing
6:22
today with regards to coming
6:24
up with drug discovery , and
6:26
all these areas are really important for
6:28
the farm and biotech industry space
6:30
. Can you point out some really high-level stuff
6:32
that's really difficult for researchers
6:34
today to achieve without this technology
6:37
that we're talking about ?
6:38
So it is basically around the
6:40
idea of using
6:42
your past data to
6:44
try to somehow make
6:46
better decisions , right ? So
6:48
at the moment , it takes around
6:50
10 years or so , and
6:53
maybe a billion or two billion
6:55
dollars to take a drug to market
6:57
. Wow , and is that the average
6:59
? That's the average . Yeah , oh
7:02
my .
7:02
God , it's not a .
7:03
Gaussian distribution , necessarily , but it's
7:05
a lot of money and a lot of time , and
7:07
this is because there's so many things
7:09
you have to do right . Drug discovery goes from
7:11
initial research of trying to
7:13
figure out what can I attack to
7:16
treat cancer , to then
7:18
trying to find agents like , for
7:20
example , small drugs that might attack
7:22
this target and might diminish
7:24
its activity , then trying
7:26
to make sure that these agents are very potent
7:29
, that they are safe for people , that
7:31
they can be ingested by people
7:33
, to then finally putting them in the
7:35
clinic and by that point you're already like five
7:37
years down , and then you have another five
7:40
years of clinical work and
7:42
then this clinical time actually 90%
7:44
of all drugs fail . So basically
7:47
you're spending five years and then
7:49
you have another five years and 90%
7:51
of the time will fail . But this kind
7:53
of influences the thinking of a company
7:55
is quite dramatically right . And
7:57
then anything that
7:59
reduces the time to
8:01
get a gold shot , which basically means
8:03
shooting into the clinic , right
8:05
. So actually trying clinic once or
8:08
improves the outcome in the clinic
8:10
is super , super valuable . And
8:13
this is what it converges
8:15
to right , because on one end , you
8:17
can use machine learning and artificial intelligence
8:20
to , for example , take your past
8:22
data on drug compounds
8:24
that you tested against the target
8:26
or something like this , and then basically
8:29
try to predict other compounds that might even be
8:31
more potent , instead of trying
8:33
all these compounds . Right , so that obviously
8:35
speeds you up because you wouldn't have
8:37
to do so many experiments , because testing
8:39
sometimes an individual drug
8:41
, you have to first make it , test
8:44
it , and so on . It might take two
8:46
, three weeks or so . So testing tens of
8:48
thousands is really not possible , whereas
8:50
you can maybe test a thousand over a few
8:52
years , and then you really have to be smart which
8:55
ones you test . That's one way in
8:57
which these algorithms can help . The
8:59
other way is , of course , once they
9:01
go into the clinic , you have very little control
9:03
, right ? You do not know what
9:05
is going to happen . However , many
9:08
companies already generated some data
9:11
on clinical outcomes of drug
9:13
compound because they have tested quite
9:15
a few over the last hundred years . Right , see
9:18
, and this is another way you can actually
9:20
improve this you can try to predict what would happen
9:22
in the clinic given a certain drug , but
9:25
there would really have very small data
9:27
and also , in the other case
9:29
, these data set are very small . You
9:31
really have to leverage
9:33
a lot of tricks from very
9:36
nice machine learning technologies to make these
9:38
things work and actually pick up what happens in the
9:40
clinic is much harder than just trying to
9:42
predict stuff that happens before the clinic
9:44
. So I think we're now this day we can try to accelerate
9:47
the work before the
9:49
clinic or preclinical drug discovery
9:51
work . What the stage of trying to
9:53
improve clinical hit
9:56
rates .
9:56
That's still very difficult and
9:58
the companies that were successful also were
10:00
more towards the preclinical
10:03
work there's a direct
10:05
link between them spending all this
10:07
time , all this money and
10:09
still having just At
10:12
best ten percent chance of getting
10:14
a home run . All that is
10:16
still linked to higher
10:18
drug prices because they need
10:21
to recoup those cost that they
10:23
put for ten years and on top of actually
10:25
reducing costs for farmer .
10:27
That itself wouldn't really help because
10:29
it could still charge whatever they want . But if
10:31
you increase competition between companies
10:33
as well by actually enabling
10:35
everyone to get to the
10:37
same level of capability , that's
10:39
also key for this and in
10:41
a way , that is also the thing we're trying to
10:43
achieve a reliving the plane of these
10:46
kind of technologies .
10:47
A large part of what's in the
10:49
medical device industry , farm industry
10:51
and biotech are Improvements
10:55
or iterations from
10:57
what was created in the nineteen sixties
10:59
, of the seventies . And
11:01
you smile over there . I was so
11:03
surprised because I remember my first
11:05
day at walking with an
11:07
abit and they were like , okay
11:09
, this is not like rocket
11:12
science . The pacemaker has remained the
11:14
same from nineteen sixty to now , largely
11:17
. But I'd like to get
11:19
your take on that statement
11:21
that I made . It might not be completely
11:24
true , but it seems like there is a certain level
11:26
of bias that sticks
11:28
around that is very
11:31
hard to think Completely
11:33
out of the box , because this is tried
11:35
and tested and has been used
11:38
with the public . You have public health data
11:40
for forty , fifty years . So you're like , okay
11:42
, you don't want to deviate from that
11:44
normal .
11:45
It's difficult to deviate from the normal
11:47
if you have to , in the end , get fda
11:49
approval , and it's easier Get
11:52
approval for something that is similar to
11:54
something that is known . But
11:56
that's , of course , only one Issue
11:58
. I think this ties very well
12:00
into what we're mainly doing
12:02
right now , which is so , first way
12:04
people build drugs . So what's more ? Molecules
12:07
, which is like tiny molecules that easily
12:09
go into cells . They have been used
12:11
. There were the first drug that have
12:13
been used for the years and farm
12:16
company is developed quite
12:18
a few of them over the last century or so , and
12:21
initially these
12:23
were found mainly from natural
12:26
analog . So people are just looking at plants
12:29
trying to figure out what the ingredient
12:31
was that you , just you are the certain
12:33
disease of cured headache . Then
12:36
people start deviating from that to be
12:38
to try to find analogs of these
12:40
like slight changes . So you can think of
12:42
this almost . As this is huge desert of
12:44
molecules , most of them are toxic to people
12:46
. There's a few islands in there
12:49
that have promising molecules
12:51
, but the desert in between
12:53
is big and vast , so
12:55
finding these islands is extremely
12:57
difficult . That's why you Tend to often
12:59
go from natural analogs
13:01
right then that
13:03
was the status quo . Then
13:06
in the 80s and 90s people started doing
13:08
more with high content screening , so they would just
13:10
generate millions
13:12
of different compounds and just screen
13:14
them over some
13:16
sort of target that they wanted to attack . That
13:19
of course gave many more results , but often
13:21
these compounds in the end tend not
13:23
to be great in the clinic Because
13:25
the experiments that people did on these compounds
13:28
were actually not very predictive of
13:30
what happened later on in the clinic
13:32
. So it was a bit more difficult . And
13:35
then , especially now after
13:38
covid , we see quite a
13:40
lot of development , in particular people
13:43
called advanced therapies , which
13:45
are like therapies based not on small
13:47
molecules but based on RNA , cell
13:50
therapy , antibody therapy
13:52
, peptide therapy and so on . So
13:54
suddenly this whole zoo of new options
13:56
that pops up and
13:58
the whole space became way
14:01
more complicated . And
14:03
also people are going back to the small
14:05
molecules again and realize , oh shit
14:07
, the ones that we actually work with we're very
14:09
similar with each other , like all the different ones
14:11
that pass through the FDA , but
14:13
often very slight modifications just
14:16
of the same molecule
14:18
. And then a company such
14:20
as in silico , for example , come around the
14:22
corner and start designing molecules
14:25
that To see . If this weird
14:27
because they haven't really been done before , but
14:29
clearly they work . So there is some
14:31
, in some ways , you can harness , like these
14:33
generative approaches to
14:35
come up with new ideas , new
14:38
solutions to all the problems
14:40
, and it's a bit like a An explosion
14:42
of ideas that's happening right now , of
14:45
different viewpoints you can apply to all
14:47
solutions again .
14:49
So it's extremely exciting , can you provide
14:51
customer success stories or something
14:53
where you are able to just take some
14:55
sample data and was able
14:58
to Improve an influence
15:00
there .
15:01
research we did a lot of consulting
15:03
work prior to actually embarking on
15:05
product development , and
15:08
some of the consulting work , for example , was
15:11
image analysis , where we
15:13
essentially help people interpret
15:15
microscopy data In
15:17
early drug R&D . And
15:19
there it's straightforward because typically it
15:21
takes people a few minutes to analyze
15:24
an image where it takes an AI working
15:26
a few seconds . So you're really talking
15:28
about For the
15:30
fifties speed up here , so
15:33
that's very straightforward . In
15:35
in other ways , we did some work
15:37
on optimizing molecules
15:39
, as I mentioned , and there
15:41
, as you can see , with some of the bigger
15:43
companies , you can speed these things
15:45
up by what five to ten
15:47
X like you can find the most
15:49
optimal drug candidate about five to
15:51
ten times faster if you use some sort
15:54
of AI powers decision
15:56
making and In
15:58
turn and this is also
16:00
where we see our initial
16:03
product now , which we have been developing
16:05
last year , because people have been
16:07
working with small molecule
16:09
in the dawn of time , right
16:11
, and this is also the
16:13
modality that most people have experience with
16:16
. So we now pushing out
16:18
or initial product exclusively for small
16:20
molecule drug discovery , where
16:23
we essentially take customer data
16:25
. Let's say somebody tested twenty
16:28
, fifty compounds in the lab , then
16:30
we would they upload the data
16:32
to our app and then they can predict
16:35
the same properties that they measured
16:37
for these compounds for ten , twenty
16:39
hundred thousand other compounds . Then prioritize
16:42
the next step , take that back
16:44
to the lab , get the results back into
16:46
the app , then close the design
16:48
, may , test and analyze cycle
16:50
and this is really what
16:52
we're doing . So it's this idea of laboratory
16:55
optimization , by reducing the
16:57
amount of work you have to do and suggesting
16:59
the work you should do next , and there
17:01
we did a few case studies , which you can
17:03
also find the no web page , where
17:05
we think it can speed up by
17:07
about four , x and it's current state
17:09
. But we're still working on some benefits
17:12
there as well . Yes , really
17:14
exciting , because you go , you design a
17:17
compound in the app which is similar to your compounds
17:19
. You potentially make it
17:21
, then you test it , then you go back
17:23
again and design the next one based on the
17:25
information you have before . Yeah
17:28
, instead of like sometimes stabbing
17:31
a bit in the dark , you can really use this
17:33
to guide your thinking and
17:35
also , sometimes something might come up which you would not
17:37
have thought about . We have molecule
17:39
that maybe you have never really synthesized
17:41
before , but make sense and then
17:43
you might find something which is maybe a new
17:45
way , says this past desert .
17:48
Very interesting . So the
17:50
way you're describing it , it's like an application
17:52
that researchers can
17:54
basically log onto on
17:56
a browser , get into
17:59
the are , into your portal and
18:01
input their data set . And is it
18:03
something that they get right away ? Or
18:05
is it one of those things where , okay , they input it
18:07
and then they head back and then you're
18:09
doing the post processing for a day or
18:12
two and then they come back , see all okay
18:14
, what these are the potential outcomes , and
18:16
then test it out . Is that the workflow
18:18
that you're seeing ?
18:20
It has to be faster
18:22
than that because otherwise , okay
18:24
, becomes more of a consulting
18:27
business again and we really want to make
18:29
these inside available
18:32
like almost instantly , because
18:34
in the end people do
18:37
a lot right . They might try this and
18:39
they might try this . They want to play a bit
18:41
around on the app . So , yes , it's a cloud
18:43
, simple clouds , vertical
18:45
SAS app you would call it , and
18:48
we initially when
18:50
an interface with which people log in
18:52
the upload a bit of data that they click a button
18:54
, predicts or other data
18:56
that they might have provided themselves , what
18:59
generated using generate a guy
19:01
on our platform and
19:03
then it should take the pain . On
19:05
the data set size , anything
19:08
between four , what's
19:10
a two minutes to a
19:12
few hours for the really big ones
19:14
. But then nobody in
19:16
early drug discovery has big data sets
19:18
because you would start with Maybe
19:21
tens of compounds , that a couple hundred
19:23
compounds . Once you have a thousand compounds
19:25
, you either failed or
19:28
one in the clinic already , but that
19:30
you don't need a I am anymore because
19:32
you already had your shot .
19:34
That's really cool the people who are
19:36
ultimately their researchers .
19:37
And on top of the game , they
19:40
come into your platform knowing maybe
19:42
one of these hundred might work , so
19:44
they've already narrowed the problem space
19:46
and then finding the oasis
19:48
in that Problem
19:51
space is a lot quicker
19:53
and probably gets more signal yes
19:55
, so basically , initially it's
19:57
just about finding something that might
20:00
what we just call
20:02
the hit , which is like maybe one
20:04
, two , three compounds , so , and
20:06
then , once you have a hit , try
20:08
to generate a few compounds that are like in
20:11
this molecule space , to just have
20:13
a look , and then at that point
20:15
we , like already have narrowed
20:17
into the space a little bit . I
20:19
really want to look into that region , which
20:21
could still be massive right , and
20:23
then generate a few in that area
20:26
. Then it's really where I was , so where
20:28
would come in ? because it would basically take
20:30
the initial Results and
20:32
then go from that to generate
20:34
new suggestions . That might make
20:36
a lot of sense , because
20:38
going really from zero
20:41
to something here Is
20:43
it would be really interesting . But
20:45
in terms of what ML AI is
20:47
capable these days it's
20:50
really not that straightforward because
20:52
there's
20:55
basically infinite molecules you can make
20:57
right and then trying to
20:59
do that . We have a grant
21:01
now pending . That tries to do something like
21:03
this is a bit more , let's say , high
21:05
risk , so let's see if it works out
21:08
.
21:08
But yeah , maybe we'll know
21:10
next year's like a baseline
21:12
, and then you walk through it and you paper it
21:14
. Now you can only do a .
21:16
I am that's awesome once you know something
21:18
and where you get that something from
21:20
is really key .
21:22
Does it get better as you onboard
21:24
more customers onto the platform , or
21:27
are these companies like super touchy
21:29
about their data sets that you can never
21:31
use everybody else's data
21:33
sets to establish baselines for
21:35
new customers ?
21:37
Let's depend . some customers are really interested
21:39
in that , because particularly
21:42
academic customers are very interested
21:45
in making data more widely available
21:47
so that sometimes even ask Whether
21:49
this is an , and of course
21:52
it can be done because our database can
21:54
distinguish between , say
21:56
, hi Lee , it's the
21:58
tiny secret , just part of the database
22:01
where very happy , happy and I
22:03
could never be used for anything but serving
22:06
to a customer . But
22:08
other customers can also . We call
22:10
centers Multiplayer . They can
22:12
engage in multiplayer mode , which basically
22:15
means that they provide some data and then
22:17
that data might be used For other
22:19
people as well , which is not
22:21
so interesting when you try to make
22:23
drugs against the particular
22:25
target , because it really tends to work on
22:27
different things anyways . But it becomes
22:30
really interesting for the things
22:32
I mentioned before the idea
22:34
of trying to predict what happens in the clinic , because
22:37
if you have , if you have a few
22:39
compounds and you know whether these
22:41
metabolize well or whether these
22:43
are toxic , group
22:45
that with many other drugs of which
22:48
you know how they metabolize or how toxic
22:50
they are , that's really cool and
22:52
there's not that much data on
22:54
that right now . If people are willing to
22:56
share that , that's
23:07
something people often willing
23:10
to engage with . It also helps
23:12
the problem , because typically people do not want
23:14
to share the structure of a compound because that's
23:16
IP , because as soon as that becomes public knowledge
23:18
, it can be patented anymore
23:21
. But if you just say , all
23:23
we can use that model to build better
23:25
models for everyone , but your structure will never
23:27
be retrieval from that model , that's
23:29
really something some people are interested in so
23:32
, max , the biggest news , at least
23:34
in the US , is ozampic and
23:36
those weight loss drugs .
23:38
It's like the biggest news . Everyone's always hang
23:41
on , so now I can . It's
23:43
a really cool thing because you lose the weight
23:45
and then it's easier to control it then . But
23:47
it's interesting that it
23:50
stems from diabetes
23:54
drugs that we've used all this
23:56
one . When I was reading through
23:58
your blogs and the papers , it
24:00
just feels hey , hang on . Something like this where
24:03
you could expand what a
24:05
drug could do or just think of
24:08
it as a new way , like a new paradigm
24:10
through an AI
24:12
, would be a lot faster than 60
24:15
years of trial over the
24:17
world's population . So that seems
24:20
like something that we
24:22
might just start uncovering , where we use traditional
24:24
drugs and running
24:27
the co-pilot and we're like hang on , this
24:29
could also do this and go from there .
24:31
Yeah , Things like we are doing
24:33
is , let's say , you have a bit of information
24:35
, initially like again 20 , 50
24:37
compounds and so which are completely new and
24:39
have never been done before , but they're not
24:41
really that potent . They're quite toxic , so
24:44
they're quite annoying , but you
24:46
can use the information from them
24:48
to basically go and
24:50
screen public databases
24:52
. There's millions of compounds out there that you can just buy
24:54
, so you can then basically use
24:56
this information to try to repurpose one
24:59
of them . For example , I don't know
25:01
, these 50 compounds are really
25:03
good against killing this
25:05
particular cancer , but
25:07
they are also doing all kinds
25:09
of other bad stuff . So maybe I
25:11
use this information from these compounds
25:13
to then screen other compounds which
25:15
then flag up and maybe these
25:17
compounds have never been used against this
25:20
disease , and then you can
25:22
start just repurposing them because they're already
25:24
FDA proof . So that's
25:26
pretty cool and history is littered
25:28
with these kind of , let's say
25:30
, super versatile compounds
25:33
, right ? Doxycycline is one of them . I
25:35
tend to discover a new application of that
25:37
every month by chance . Then
25:40
Viagra is actually one of the most
25:42
famous ones , because Viagra was
25:44
initially . I can't remember what the
25:46
clinical trial was about Hard
25:48
, hard .
25:49
It was the hard .
25:50
Yeah .
25:51
There's also so many drugs that kind of failed
25:53
clinical trials , but they were already
25:55
shown to be safe for
25:57
patients and we don't even know
25:59
if they can do anything yet . So
26:02
that's also then really sad and often . These
26:04
then sometimes become available
26:06
and you can maybe ask a farmer company
26:08
if you can look into them on top of
26:10
this vast chemical space and say
26:12
we haven't yet explored . We still don't
26:14
really fully understand how , first of all , some
26:16
of the drugs we use actually work and
26:19
then , second of all , how
26:21
we can actually repurpose them to
26:23
do even completely different things , right
26:25
, because we already know that they're safe . Right , so
26:28
they can do so many things
26:30
potentially . It was really exciting . We
26:32
have medicine . We only scratch the surface
26:34
until now , and now we like
26:36
starting developing all these new therapies , all
26:39
these different approaches . Cell and gene therapies
26:42
are really awesome , especially for like cancer
26:44
.
26:44
Yeah , yeah , it's super exciting .
26:46
One of the things we wanted to ask was as
26:48
you envision the future of deep
26:51
mirror and the problems that you're solving today
26:53
. You look five , 10 years down
26:55
the line . How do you see
26:57
that ? What are the challenges that you
27:00
foresee right now that you're solving
27:02
, and how you'll get to the future state that
27:04
you want ?
27:06
The biggest challenge for any , let's say
27:08
, software company in this space has
27:10
always been to
27:12
resist digging for gold
27:14
, almost . Because once you have
27:17
something that kind of works
27:19
, everybody will try to tell you oh
27:21
, now you have to go all the end to the
27:23
patient . But then you tend to put
27:26
all your money onto a single component
27:28
at some point and then again
27:30
90% failed and then
27:32
you might have your cool platform software
27:35
which sped up your preclinical
27:37
work . Let's say it's five , x or
27:39
something like that . But that still
27:41
doesn't necessarily mean that afterwards you
27:44
will have an higher success rate necessarily
27:46
. But people still
27:49
think in this old school biotech
27:51
way where you basically go and
27:53
okay , so you play around a bit , you find
27:55
something , and then you put all your hosts and then you
27:57
become a proper biotech company that
28:00
then gets bought by a farmer or your partner
28:02
with some farmer companies and
28:04
so on , and staying clear of that
28:06
like trying to resist this digging
28:08
and really trying to stay a service provider
28:11
and delivering value to everyone
28:13
. That will be quite tricky
28:15
, but we see other companies in this space
28:18
that kind of manage to become
28:20
a fully vertical Sastau
28:22
solution . Benchling is maybe
28:24
one of the most famous one . Nobody
28:26
thought that something like Benchling would work when they came
28:29
around because effectively
28:31
it was like a workflow solution for researchers
28:34
, and why would researchers pay for that ? But
28:36
then over time people realized researchers go to
28:38
farmer companies , have big pockets and
28:40
suddenly Benchling has contracts that are more
28:43
than a million and annual revenue .
28:45
You said something that is so interesting
28:47
. I
28:49
see the promise of
28:51
this technology . I'm thinking , oh
28:53
, massive amounts of data . You put AI there
28:55
. We will be able to uncover
28:57
certain patterns . But you mentioned
28:59
something so interesting where sometimes
29:02
it might just be helpful to get
29:04
a cleaner set of data , a smaller set
29:06
of data , and then expand from
29:08
there . So that is really
29:10
interesting I have not thought of that
29:12
before and how AI can help with
29:15
.
29:15
Sometimes the less is more You're also
29:17
defining the question very well
29:20
for that particular data set . Like
29:22
in our early days , we once got contacted
29:25
by a company that essentially asked
29:27
us oh , we have these millions
29:30
of images here . Can we pipe
29:32
those through an algorithm and learn something
29:35
? And then again
29:38
, initially you think , OK , maybe
29:40
, but then you look into this and then
29:42
, as long as you don't really ask a
29:45
proper question about the data , you
29:48
just end up with nothing , Like
29:51
if you don't ask a question you don't know . you
29:53
won't find an answer because everything will be statistically
29:55
significant , and then you might be
29:57
amplifying noise as well , because it's a big data
30:00
set and you can't just have a look at it , and
30:02
you might not learn anything . It's like this whole
30:04
idea of looking at each grain
30:06
of sand on a beach and then trying to understand
30:08
what a beach is . It doesn't
30:11
really help . You have to have some sort of
30:13
question about the data you might
30:15
analyze . There's something similar
30:17
happening in academia now , where people were
30:19
thinking a lot about something called spatial
30:22
transcriptomics , which is basically the idea
30:24
of taking images of tissues
30:27
and then also looking at genetic information
30:29
in these images and basically generating
30:31
not even terabytes but petabytes
30:34
of data on this . And still
30:36
they don't really know how to apply this , because
30:38
just by looking and collecting data
30:40
without narrowing the question
30:42
, we don't know . This is where
30:44
we come in , right that people still are very
30:47
much required to define a
30:49
question , define the hypothesis
30:51
and then help get a
30:53
machine to help with the number crunching . Right
30:55
Can't replace it yet .
30:57
Yeah , max this was really
31:00
great . Before you leave us , we always
31:02
like to give our guests the
31:04
mic . You already have a mic
31:06
, but we want to pass the virtual mic
31:08
to you so that you can give a shout out to your team
31:11
, the work that you're doing and where people
31:13
can reach you , because we have a few
31:15
OVCs and founders who listen to
31:17
us and they're always interested to hear
31:19
what people are building and how they
31:21
can reach them .
31:22
Yeah , people can reach me straight on the max
31:25
at deepmirrorai , and
31:28
without the team , nothing of this would
31:30
have been possible . I'm
31:32
quite fortunate because I have two co-founders One
31:35
is more on the technical side , one is more on the product
31:37
side . I think the
31:39
synergy between the two of them is pretty
31:41
much amazing . We
31:44
also over the last year hired two
31:46
more people who help us more with
31:48
customer service and things , like Cecilia
31:50
and Jacob
31:52
, who's like basically
31:55
taking charge of all the machine learning , and potentially
31:57
in the next year we'll grow the team to
31:59
10 people . So I think
32:01
we might be looking to hire soon . That's
32:03
very exciting and at the moment we're very much
32:06
bootstrapped for our consultancy
32:08
contract . We have a bit of investment
32:10
. We might be looking for some later
32:12
this year to re-scale up the product Nice
32:15
. But , yeah , that's all . Very
32:17
. Sometimes I would say we're the pre-seed just
32:19
looking to find the product with which we'll go
32:21
on the seed rocket . But
32:24
you really have to . In
32:26
such a complex space , as we mentioned , you really
32:28
have to watch out that the thing you're building is actually
32:30
something useful .
32:33
I wanted to say thank you for coming on the show . It
32:35
was really nice to meet you and hopefully we
32:37
have you on again . Check in a couple
32:39
of years' time . See how deep mirrors do we
32:45
leave you with thought . The
32:47
future of the pharmaceutical industry
32:50
is on the brink of transformation
32:52
. From AI-powered drug
32:54
discovery to the emergence of novel
32:56
therapies , the landscape is
32:58
shifting rapidly . These
33:01
innovations usher in a new
33:03
era of health care . Only time
33:05
will tell . Stay tuned , stay
33:08
informed and join us again to
33:10
explore how things are changing
33:13
around the world . Until
33:15
next time , stay curious
33:17
.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More