Model Medicines AI Drug Development

“We didn’t start Model to launch a company — we started it because no one else met our standards,” says Dr. Daniel Haders II, CEO and founder of Model Medicines. On this episode of Vanguards of Health Care, Haders speaks with Bloomberg Intelligence analyst Andrew Galler about building an AI-native drug discovery engine capable of identifying cryptic binding pockets and designing first-in-class drugs. They unpack why hit rates and novelty must go hand-in-hand, an overview of the company’s pan-antiviral MDL-001, and why generalizability — not reinforcement learning — is the true litmus test for AI in biotech.

See omnystudio.com/listener for privacy information.

Episode Transcript

Speaker 1 (00:21):
Hi everyone, I'm Andrew Galler and welcome to another episode
of the Vanguards of Healthcare podcast. I'm one of the
Bloomberg Intelligence analysts covering biotechnology. Bloomberg Intelligence is the corporate
research arm at Bloomberg, encompassing five hundred analysts worldwide. I'm
there to be joined today by doctor Daniel Hayters, a
second CEO and founder of Model Medicines. Daniel, thanks for
being here today.

Speaker 2 (00:42):
Thank you so much for having me.

Speaker 1 (00:43):
So just to kick it off, maybe with something topical
and recent a thought piece published by the Model team.
You recently drew in a quantitative AI paper published by
your team, and a quantum and ants drug discovery paper
by Adsilico declaring twenty twenty five the year of hybrid
drug discovery. Maybe just touch on you view the individual
strengths of the two approaches and then how they synergize.

Speaker 2 (01:04):
Sure, So, I think the biggest thing that I'll start
with is what makes us similar, and then we'll go
to what makes us different. So I think what makes
us similar that we were calling out in that piece
was an idea that the future of the pharmaceutical sector,

(01:25):
especially as applied to AI enhanced or AI driven drug
discovery will be in fact driven by drug discovery, and
in serious, detailed drug discovery, which is ultimately the goal
of what we're doing here. And so in that piece,

(01:47):
what we first bring out is that true drug discovery
means that success is measured first by in vitro validation,
then in vivo validation, and then clinical validation. But not
just of compounds or basic chemistry, but of drugs. And
of course there's a difference between a compound and a drug.

(02:09):
A drug is unique not just from a USPTO point
of view, but also a mathematical point of view. We
use a technique called Tanemoto scoring. It's also a drug
that's safe and well tolerated. It's one that is dosed
a way that's appropriate for the disease at hand. In
many cases that's oral, and dosing is regulated by only

(02:33):
being saved once a day, ideally for many indications. And
so those are the things that make us similar. We
both hold ourselves to that standard that we want to
build drugs. We happen to use AI to do that
along the way. Now, there are some differences in approaches
that we take. I won't of course speak in detail

(02:56):
for in Silico. But their process recently, they've been using
a quantum computing approach, and in that they were evaluating
millions of compounds. In our case, we were using a
generative AI approach, and there we were able to create
a chemical space map of over fifty trillion novel compounds

(03:21):
and then allow our model to search that fifty three
trillion compound chemical space to be able to arrive at
twelve compounds ultimately nominated for synthesis and evaluation pre Clinically,
I'll note that all twelve of those worked exactly as predicted.
But even there where we went one hundred percent in

(03:44):
our hit rate, right, twelve for twelve in terms of
compounds with therapeutic concentrations at which they were effective. You
may contrast that to what they're doing. They're doing something
very you know, a little bit different. It's a smaller
solution set that they're searching from. But their and their
hit rate may be a little bit lower. Yes, but

(04:07):
in terms of did they discover a full and complete drug,
which appears to be yes as well. Right, And this
sort of speaks to back to this point. We're able
to search very large spaces. We're able to get exceptionally
high hit rates to discover libraries where we go twelve
for twelve in terms of discovering compounds with effication that

(04:30):
are efficacious that therapeutically relevant concentrations. But at the end
of the day, it's not only about the twelve, it's
about going It's about finding the one drug that solves
that disease for that patient at that moment, and that's
more important than anything else.

Speaker 1 (04:49):
So now it's theer in on your AI platform A
little bit. Model seems to be taking a more aggressive
approach than some of your peers. What I mean by
that is, while many companies are talking taking a sort
of low hanging fruit approach to validate their program first
with a known target, maybe a few tweaks on a
known molecular modality, model is going after a novel protein
target and a sort of holy grail indication of a

(05:11):
broad spectrum antiviral. What was it thinking behind that?

Speaker 2 (05:16):
Yeah. To understand why we took that aggressive approach, I
think you need to step back to why we launched
the company, and it's important to remember we were not
we are a group of scientists and drug developers. But
when we started thinking about this space, a lot of
the core team was approaching it as investors. In fact,

(05:38):
we were looking to make an investment in this space,
so as as disciplined investors, we first built a set
of underwriting standards that we were looking to write a
deal to. We spoke with a lot of companies, larger
ones who had already launched and were down the road,

(05:58):
and a lot of emerging start up and we ultimately
did not see a company that fit our underwriting standards.
And those underwriting standards were based around a few key observations,
namely that there are an exceptional amount of a very
talented medicinal chemist and biologists that populate academic laboratories and

(06:23):
pharmaceutical research laboratories, and those scientists are very good at
doing doing some of the things that you were mentioning right,
some of those less aggressive things where we have known biology,
where we have known chemistry that works, and then looking

(06:45):
to build upon that. When we thought about applying artificial
intelligence and machine learning to drug development, we wanted to
build a company that leveraged AI to do things only
AI could do, and namely, there we wanted to discover
new pockets and not just any new druggable pockets, but

(07:08):
for instance, in our infectious disease program, ultimately discovering a
cryptic alisteric pocket right that exists across a significant portion
of RNA viral biology, and that's something. Cryptic pockets, for
the most part, are something that have only been discovered accidentally,
never on purpose. Their discovery has never been engineered. We

(07:32):
also wanted to focus on drugging the undruggable, so going
places where the pharmaceutical industry and human scientists had attempted
to but had been unsuccessful in discovering drugs that alleviated
disease via a target. And this is where we thought
we should be. And to say this a different way,

(07:54):
we thought that AI drug discovery companies should only focus
on the discovery of first in class therapeutics via novel
mechanisms of action or that overcome undruggable targets to bring
a first drug to market with against that target, and
they should set new standards of care. And of course,

(08:17):
inevitably as you go out in time, maybe it doesn't
work out that way every single time, but we believe
that should be the ambition. The discovery of first in
class therapeutics via undraggable targets or the discovery of new
targets every single time. And if we do that, we
can finally truly make it dent in human health via

(08:39):
the discovery of therapeutics in a way that really hasn't
occurred to date.

Speaker 1 (08:46):
And you kind of were touching on this. As we
think about the explored chemical space and what a small
proportion of these pharmacologically active molecules, which is anywhere from
ten to the twenty third to ten to the sixtieth
depending on who you ask, or such a small percentage
of this, how can generative AI or other AI models
help to expand our explored space when it's really just

(09:07):
train on existing molecules. Is it synthetic data, broader inferencing?
What do you think it is?

Speaker 2 (09:14):
Yeah, so this is a really good question and at
the core of what we do and what the industry
is doing, and it really is about extrapolation. To your point,
there's somewhere chemical space is somewhere between the ten of
the twenties and the ten of the sixtieth. But when

(09:37):
we compare so, what we don't understand, or what most
people don't fully understand, is how much of chemical space
has currently been traversed. So all the drugs that have
ever been in some form of clinical trials or the
equivalent is on the order of ten thousand compounds. So
whether true chemical space is ten to the twenty fifth

(09:57):
or it's ten to the sixtieth, either one of those
numbers when you subtract ten thousand, is still that number, right,
And so in reality, if chemical space is ten to
twenty fifth or ten to the sixtieth, essentially all of
it is really unexplored at this point. So how do
we overcome that? So there are some folks that think

(10:20):
about this from the point of view of, well, if
I can get more data from wet lab experiments, then
I can get to deeper portions of chemical space. And
we don't believe that that is untrue. But we also
don't believe that that is properly using artificial intelligence either.

(10:45):
We think that the promise of artificial intelligence is not
simply create ten to the twenty fifth compounds and test
them so that that then becomes the training set. We
believe that the promise of AI is that you can
essentially use the data that we already have build better models,

(11:06):
build better data pipelines, build more stringent training environments and
then create models that are able to extrapolate into that
deep chemical space. Right. And so I think this is
a fundamental difference for model medicines. Rather than us physically
creating compounds that sample the entirety of chemical space and

(11:29):
using that to train models. We believe that there's approximately
enough data already created if you build models the right
way to extrapolate into that deep chemical space. So this
is a fundamental difference that we have. But it is
model extrapolation. I believe that is the most important feature

(11:54):
of AI drug discovery, right, That is where the discoveries are.
It's the ability too, if you will, if you're thinking
about a four quadrant graph, right, it's the ability to
train on data in the lower left hand corner of
that of that four quadrant map and be able to
discover compounds in the upper right hand quadrant quadrant of

(12:18):
that chemical space map. Right. It's not a it's not
a uh uh. It's not necessarily a feat to train
on upper right and discover upper right, or train on
on lower left and discover lower left. The feat is
picking one corner and training on it and then discovering
in another corner, and that's what we set out to do.

Speaker 1 (12:41):
It sounds like you think generalizeability is the most important
quality for a good AI model. Is that there to.

Speaker 2 (12:47):
Say, yes, generalizability, yes, But but I also do not
necessarily believe that you have to have a single a
single model that is capable, a single model train once
that works for all targets in totality. Right. I think

(13:11):
it's possible to come in with that sort of idea
that foundational models are important. But again, I'll go back
to this idea. Anytime you're doing drug development, once you
have identified a target that you have then validated, you're
just discovering chemistry for that one target that has all

(13:34):
of the ADME and safety talks features that you need
for it to be successful. And so I don't necessarily
think that the creation of generalizable models in itself has
to be the endpoint, right, a single model that can
go everywhere, because ultimately we're discovering drugs for specific targets.

(13:56):
But I do believe you need to have models that
can generalize as well. So that's it's a it's a
nuanced take potentially, but one that I believe is very
important that I'm happy to get into a more detail
if you like.

Speaker 1 (14:12):
Yeah, absolutely, I mean, I definitely think the AI field
right now is kind of lacking in nuance when you
look at the popular literature on the space. But I'd
like to actually dig it into a bit on what
you talked about with data pipeline, because when I think
about Galileo, your AI platform and the data sets you
do with the purpose build data pipelines, you've toauted things
like fifteen increases and key starve bioactivities, among other increases

(14:37):
relative to competitor data sets. So ye, to an extent,
like more data and richer data is important to train
these models.

Speaker 2 (14:46):
Yes, yeah, yeah, And that's why So this is an
excellent point and why I think the nuance this is
where the nuance is important, is that, for instance, one
of the things that we discovered along the way to
our discoveries is that, for instance, concepts of implicit data

(15:09):
and explicit data. For instance, so what we discovered in
our infectious disease work is that there's quite a bit
of explicit data where you have you have a compound,
you have some type of black box or assay, and
then you have a result, right, and that can be

(15:32):
as simple as a compound and a protein and a result.
It can be a compound and a human being and
a safety read out. It could be everything that's in
between there. And what we found is that there is
a decent but limited amount of explicit data where you

(15:53):
explicitly know how the chemistry is connected to a specific
biological result. What we then discovered actually is that we
could use AI to discover data that is technically in
the public sphere, but for which people don't know how

(16:14):
the chemistry and the result are connected, right. And so
what you see is there's been quite a bit of
work that's done in high throughput screening, for instance, or
an animal model training data, where you might take fifteen
thousand compounds in a library, you run them on a
cellular model of disease, and you get a result and

(16:38):
the drug either alleviates the disease or it doesn't, but
you're not sure how the drug is alleviating the disease
or not alleviating the disease. We call this implicit data.
One of the things we do in our data pipeline
is that, yes, we use the explicit data. We ingest
all explicit data in the same way and open AI

(17:01):
ingest data. But then also what we do is we
use AI to essentially what I'll say are de encryption
keys for all of the implicit data that is available
in the literature for which we have a chemistry and
a result, but we don't know how that result is
mediated by that chemistry. And so when we do that,

(17:23):
we're able to create multiple x more data. Oftentimes it's
the most important data and the data that's key to
novel discovery from that utilization of both implicit and explicit
as well. One other thing I'll point to as well,
and is how how you split your data right, And

(17:48):
this is really important to extrapolation or generalizability. And what
we find, I believe too often is that the data
that we are training on is oftentimes too close in
chemical space to the test sets that we're evaluating the

(18:10):
efficacy of the model on. And what we've found, counterintuitively
is that if you create more stringent training environments, your
retrospective validation of your model will look worse, but your
prospective drug discovery potential of your model will work better.

(18:34):
And so this is different than maybe the dogma that's
developed where you want to create a perfect model retrospectively
to create the best potential model prospectively, we actually found
by training that same model harder, getting a worse result retrospectively,
you get a better result prospectively. So I bring this

(18:54):
all up to say that I often i'm on panels
or I talk to folks young scientists, and they or
or investors, and they say, what's the secret sauce? And
the truth of the matter is is that there is
no single secret sauce. There's no one thing that we
do to get the results that we've gotten to this point.

(19:17):
There are a one thousand point solutions that we've adapted,
whether that is the creation of these encryption keys for
implicit data, whether it's creating more stringent data pipelines or
training training environments, and all of these things add up
to models that are able to extrapolate and models that

(19:40):
are able to create truly novel compounds that solve disease. Yeah.

Speaker 1 (19:46):
Absolutely, So I want to touch back on something you
said a little bit earlier about one You said your
model is high hit rates and low Tanamoto scores. So
it's it fair to say that, I mean, you're going
into these kind of unknown chemical spaces with high accuracy.

Speaker 2 (20:03):
Yeah, that that's a great that's a great way to
uh to say it, and it's it's a way to
raise the bar of what novelty is while raising the
bar of what should be expected from artificial intelligence. So
let me let me explain that in a little bit

(20:24):
more detail. We believe that if you are using credible
artificial intelligence and machine learning to discover therapeutics, that we
you should see gains in terms of compounds nominated by
your platform that work exactly as engineered to work or

(20:49):
predicted to work in the laboratory at exceptionally high rates. Right, so,
right now, to take a step back, high throughput screening
typically has hit rates in the neighborhood of half of
a percent, and early work in AI driven drug discovery
was low single digits or sorry, high single digit hit

(21:11):
rates to low double digit hit rates. We believe that
this still wasn't good enough and demonstrated demonstrated that models
did not fully understand what they were predicting. Right, And
part of this is wrapped up in an idea that

(21:31):
we want our models to understand how to predict drugs
that work and predict drugs that don't work at the
same time. What we control call controlling the positive chemical
or biological space as well as the negative biological chemical space.
So when we built our models, we built our models
to be able to do both of them. And that's

(21:52):
how we believe that our models. That's the threshold that
we demand as achieved of our models to demonstrate that
they truly know what they know, and when you bring
those efforts together, you get these high hit rates. So
it's the reason why we can do a one shot
drug discovery program. Have the model looked through fifty three

(22:14):
trillion compounds, have it nominate twelve, synthesize all twelve of them,
and all twelve of them work in the laboratory as
as predicted. Now, hit rates, though, are different than the
novelty of those drugs to the training set and to
each other. Right, So it's relatively easy to if I

(22:38):
were to tell you we discovered twelve compounds that all
work is predicted, but all of them are essentially identical
to each other. From this point of view at Tanemoto,
what's much more difficult is that to discover compounds that
are one novel to their training set and then two

(23:01):
novel to each other withinside the discovered solution set. And
so that's the bar that we set for ourselves. For
those deep in this industry, what our target is is
that at a minimum, our novelty of our discoveries to
the training set or at least zero point four or below,

(23:24):
and that's sort of a minimum threshold. And then novelty
on the Tanamoto scale to individual compounds is zero point
two or less. And that's what we strive for in
each and every program, such that we're not just discovering
new drugs at a high hit rate, but those drugs

(23:45):
are novel to the training set and to each other.
And what that allows us to do is that gives us,
in this case, twelve data points spread out through chemical
synthesis space, and we inevitably find that one works better
than the other. We can triple down, if you will,
on those areas of chemical space where we find novelty

(24:06):
and we find activity, to find the most novelty and
find the most activity on these novellly discovered islands of
chemical space.

Speaker 1 (24:17):
Maybe just to dig down on those low Tanamoto scores
that you implement as a domain in your models. I
think the inner Bayesian in all of us would say
there's a lot more ways to be wrong in the
unknown space. So how do you think about validation and
the risk of hallucinations when you go that far in
and go for that dissimilar of molecules.

Speaker 2 (24:36):
Yes, so this is this is a great This is
such a good question because inevitably generative AI techniques, those
those models will there will be models that produce results
that hallucinate. There's no doubt about that. And that that's

(25:00):
why when we do drug discovery, we don't leveraging our
generative AI pipeline. We don't simply allow the model to
generate compounds that we then synthesize and take into the laboratory.
What we do is that we'll oftentimes go through a

(25:22):
zero shot and a one shot discovery process. So the
way our pipeline would work is that we would create
an initial data set, an initial model, We would train
the model on that data set, it would nominate compounds
for validation, we would take those into the lab. And

(25:42):
by the way, too, I always like to make sure
everyone understands we are a dry lab company. So when
we make a discovery with our platform, we have that
synthesized by a partner and then we send that to
a partner lab for evaluation, will then bring those hits back,

(26:03):
and then we'll leverage generative AI using those hits as
seeds and will generate new chemical space around each of
those initial hits. What we then do, though, is we don't,
like I said, we don't simply synthesize those Generative AI
created compounds. Will then bring back the original model that

(26:25):
we created to traverse the entirety of the Generative AI
recommended nce compounds, and then that's how we get from
fifty three trillion back to twelve. So through the process
of using both generative AI and our original selection model,

(26:45):
we're able to put to the side any of the
compounds that are hallucinations by the platform and again focus
just on compounds that the model believes are are relevant
to the goal at hand. And so in this way,
we're using multiple we're using models on top of models,

(27:07):
and models to filter models. And this is one of
the ways that we are able to achieve the success
that we're able to achieve today. Because I guess maybe
it's set a different way. We don't assume that our
model will not hallucinate, and so because of that, we
introduce multiple models filtering one another, and so therefore we're

(27:28):
able to overcome that potential limitation.

Speaker 1 (27:32):
And then maybe as we compare your model to other
models on hit rates in particular, I think that there's
I think part of the different the disparity is maybe
a reliance on reinforcement learning for some models, like I
think back to the paper for Reinvent four for example,
and as you go from loipop one to EPOC, maybe
fifty you have I think, I think it's going a

(27:54):
fifty fold increase in hit rates granted to telling gain
to about four and a half percent. But do you
think there reliant on reinforcement learning for these models makes
them artificially depressed on the hit right.

Speaker 2 (28:05):
Yes, so so I do, and and I think we
go we can go back to certain ideas as well,
is that error never gets minimized right as you manipulate.
As you manipulate error it multiplies. And also once you're

(28:25):
down a certain pathway, the models can get stuck in
certain pathways potentially based on based on how they're set up.
And so I think that that's that's a real that's
a real concern. I think one of one of the
things that we attempt to do right, and we are

(28:47):
we we are humbly trying to constantly be better here right.
But one of the things that we're well aware of
all the time is dogma that is in science in general,
dogma that is in the literature, and then even dogma
or bias that can be unintentionally passed on to to models,

(29:13):
such that your model and the people that are in
the drug discovery program get stuck in certain areas of
chemical space, or get stuck in certain ways of thinking
or interpreting data. And so what we try to do
is set up our models such that there is no

(29:37):
existing dogma or bias that's put into them. The data
simply is what the data is, and we want the
model to try and sample various different areas of chemical space,
and not for us to assume that we know anything

(30:01):
a priory or for the model to necessarily overestimate how
much it knows a priory. So I think you're exactly right.
There is this danger of getting stuck in an area
of chemical synthesis space and not being the model, in
the people around the model executing the model not being

(30:23):
able to get out of that, and so I think
that is a real issue that occurs, and a lot
of times what you'll see as well is that this
will come together. You'll see this in the limited chemical
diversity of the hits that are discovered by certain models

(30:45):
in certain runs. Right, So, for instance, if you discover
twenty compounds and those twenty compounds lack or have significant
similarity in them, what you know is that your models
only traverse really one area of chemical space. It has
it independently gone out in multiple different directions to discover

(31:10):
multiple islands of chemical space that all may be able
to deliver a solution.

Speaker 1 (31:15):
And is it this feature of your model that kind
of led to this broad pipeline? Like when I look
at the model medicine pipeline, you have you have your
lead mdls or zero one for rd RPM one, but
you actually have twelve backup compounds for it, which is
I would say unusual for the biofarm industry. But is
it just kind of this searching for different chemical islands

(31:35):
it leads to.

Speaker 2 (31:36):
That, Yeah, that is that is exactly That is exactly
what we're doing. Now. What I like to tell you
that we're going to use a zero shot approach every
single time, and we're going to discover the drug every
single time, across potency, across safety, intolerability, the across at

(32:01):
ME and pharmacokinetics. I'd love for that to be the
way that it ultimately works, and we're engineering our products
to be that. But I think in reality, what I'm
looking for when we're doing zero shot drug discovery is
not necessarily that we're going to find the one in

(32:22):
terms of an individual compound, but we're finding novel islands
of chemical space that then we can come back, as
I was sort of mentioned before, leveraging generative AI and
the knowledge of these islands and exploring those islands to
find the idealized chemistry on a particular island, and then

(32:43):
what we ultimately want to do. If we have a
diversity of these islands, where we have activity and appropriate
at ME and safety talks features, then if you will,
these islands of chemistry can compete against one another in
the laboratory and in vitro studies and in vivo studies

(33:03):
to ultimately see which island, if you will, of chemistry
is the best. But we don't want, again coming back
to this idea, we don't want twelve compounds on one
island of chemical space. Ideally, what we would rather have
is that for the model to nominate twelve unique islands

(33:24):
of chemical space, evaluate them pre clinically, and then be
able to come back and determine these three islands, or
these five islands, or where we should play next. And
so that's what we consider truly to be interesting because
in that case, we feel like we're doing a better
job of beginning to understand what we don't understand today. Right,

(33:49):
so if we just find one island, we only were
leaving out the rest of space, where we're not even
considering what could be beyond there. In this case, when
you have islands of chemical space competing against one another,
we truly believe it's the fastest way to getting solutions
to human beings with disease. And that again, that's where

(34:12):
we always come back with is ultimately the goal.

Speaker 1 (34:15):
We've talked a lot about chemical space, but I wanted
to get your thoughts on one other piece of conventional knowledge,
so to speak, for drug development regarding only ten to
fifteen percent of human genes being druggable. You see that
expanding over time as well, whether that's the discovery of
cryptic binding pockets or something else.

Speaker 2 (34:33):
Yeah, I think there's I think there's no doubt about this,
and not to go to historical or maybe philosophical about this,
but I'm there's already been multiple times in my lifetime
where I remember people saying, well, this is it. We've
solved this, We're not going to be able to go

(34:55):
any further. I mean talking about somebody that was I'll
date my self born in the late seventies and into
the PC revolution. I remember people saying that, you know,
our semiconductor manufacturing is going to get to a certain
limitation in terms of material science, in terms of size,

(35:18):
and that's going to be it. That's going to be
the best chip we're ever going to be able to make.
And I feel like when I was a kid, people
were predicting that that was going to occur sometime in
the early two thousands, and here we are right. And
so I think I think the truth of the matter
is that today it is a truth that there is

(35:38):
only X number of drugable targets. But that's based on
the knowledge that we have today. What I truly believe
is much like the breakthroughs in the semiconductor industry, there
are truths. There are their first principles truths in bology

(36:00):
and chemistry and in biochemistry specifically that we do not
fully understand today. But once we understand those fundamental truths,
those will unlock more drugable targets. Right. So, I think
it's true that our druggable targets are limited today, but
with advances things like we're doing, things like others are doing,

(36:24):
I think that will unlock new targets that will be
accessible to platforms like ours and future platforms to be developed.

Speaker 1 (36:32):
So it seems like biologically and the chemical space, it
seems like AI is expanding the capabilities of drug development,
is fair to say in your.

Speaker 2 (36:42):
View, Yeah, I think it's expanding. It's expanding the ability
to get to the four corners of that ten to
the twenty fifth, ten to the sixtieth. But to your point,
maybe what we haven't talked about is is biology is
considered by some to be a ten to the twenty

(37:04):
six problem, right, And so there are areas of biological
space that we don't fully understand now, but once we
understand them, we will then discover that, oh, here's an
island that we didn't think of biology, that we didn't
think was druggable. Now we know that it's druggable, and

(37:25):
then very likely there's likely to be an area of
chemical space that's never yielded a drug, but that area
of non productive chemical space is likely to be applicable
for that new area of biology that we didn't know
could be drugged, you know, just a few years ago, right,
Technology development and life sciences. Technology development has a habit

(37:50):
of working this way, and I think the best way
to approach it, or at least the way that we
approach it, is that we don't assume that we know
anything before we start a project. We don't assume that
there are limitations, and then we discover limitations along the way.

(38:12):
And I think there's a lot of people like us
out there working in this way. And I think when
you work in that way, you discover things that are
unexpected by definition, and then you discover solutions to solve
those unexpected discovered problems, and then we go from there.
So I don't think this round leveraging AI and machine

(38:35):
learning in the bio and chemical space will be any different.
I think fundamental knowledge will continue to be developed that
will unlock biology, and that unlocking of biology will unlock
new chemistry and will wash rents or repeat and go
from there.

Speaker 1 (38:51):
So maybe we can move on from broad AI platforms
and as POTENTI more broadly and talk about some of
your individual programs. If that works. So MDL zero zero
one is your lead. Maybe just start off by describing
the genesis of this program.

Speaker 2 (39:04):
Yeah, I appreciate the opportunity to talk about this because
obviously this is near and dear to our heart and
also is something that enabled us, provided us the opportunity
to put some of our ideas about drug discovery and

(39:26):
drug development to work. And what happened is that we
had launched Model Medicines really as a research project in
twenty nineteen, and we went to the JP Morgan Healthcare
Conference in early twenty twenty and started to take meetings

(39:47):
and you know, started to privately show some data to
folks and indicate that we believe that we might be
onto something. And when we got home, of course, those
in the industry who come back from JP Morgan Healthcare
Week oftentimes have what people call the JPM flu. It's
a lot of early mornings and late nights, and a

(40:09):
lot of people are exhausted and are typically down for
the cow for a few days. If not a week
after the conference, but obviously in January twenty twenty, something
was different and we soon realized how different it was.
And Withinside Model Medicines, we had a conversation where we

(40:31):
determined that if we were going to be serious about
building this company, that this was exactly the type of
problem that we wanted to tackle and exactly the type
of problem that we believe was built for AI driven
drug discovery. And namely, what we did first, which is

(40:56):
what we pride ourselves of doing, is we we assumed nothing,
and we gathered together in a number of meetings world
renowned clinicians and research scientists in the virology space, people
like doctor Davey Smith of UCSD, people like doctor Chris

(41:19):
White and doctor Adolpho Garcia Hysteria at Mount Sinai and
doctor Assumat Chanda and are now Chotategy at Scripts Research,
among a number of others, and we asked them to
educate us on why we were in the situation that
we were in Q one of twenty twenty, and they

(41:42):
provided us with a history lesson and a target product profile,
and namely, they demonstrated to us that an infectious disease
drug discovery had always been done in arrears, and no
one had been able to ever build a program for
the next viral pandemic before it existed, and they said

(42:06):
that should be the goal. And they also pointed us
to this idea that the drugs needed to fit a
world population, meaning that the drugs needed to be orally
available so that it could be given equally to someone
in New York City or someone in Sub Saharan Africa

(42:27):
or the Indian subcontinent. And it needed to be a
once a day drug that didn't require special handling or refrigeration.
And ideally it would be so safe and well tolerated
that not only could it be used as a as
a therapeutic, but it could be dosed pre exposure prophylactically

(42:49):
in the event of a future pandemic. So it was
with this that we that we built a target product profile.
And on the way is we're building this, we said, well,
to go about this, we're really going to have to
discover a new target across viral biology, and it's going

(43:10):
to have to be something new, because, as correctly predicted
by this group, the drugs that would come to market
would likely be protease inhibitors. And nucleosides, which are drugs
that are ultimately can be very effective, but would never
be able to be broad spectrum anti virals, again capable

(43:30):
of protecting us against the next pandemic before it existed.
And so it was with this we took that target
product profile back to our dry lab, so to speak,
and we started building. And the first thing that we
wanted to discover was a new druggable target. And in
our search we identified an alisteric cryptic alisteric target on

(43:56):
the viral polymerase one that we found to be functionally
conserved across viral biology. And that pocket was slightly different
on different viruses in terms of there was a slight
difference in the volume of the pocket, the shape of

(44:17):
the volume, and even some of the surface chemistry within
the pocket, but nonetheless it was functionally there because the
function of that pocket dealing with conformational changes of the
plymeras that take it from an inactive state to an
active state. And so once we had latched on to
a potential target that existed across large areas of viral biology,

(44:42):
then we had to the next problem, which was the
problem of chemistry. To solve this target so, as I
mentioned before, although functionally it was the same target we
had identified, the target was really more a family of
targets than a single target that you to drug, because remember,
like I was saying, the volume was a little different,

(45:04):
the shape of the volume, how deep the pocket was
from the surface, chemistry was slightly different. So we thought
we needed to create one drug for numerous related but
different druggable pockets simultaneously. This again is something that really
only AI could do, or we believed AI could do.

(45:26):
And we set our model to the task, and it
nominated MDL zero zero one, and it believed the model
believed that MDL zero zero one represented what I'll call
a global minimum energy if you will to borrow that term,
that would allow us to drug multiple viruses simultaneously. We then,

(45:48):
working with partners at SCRIPTS and Mount Sinai Utah State University,
as well as with NIH, tested that drug mdls erser
one across a battery of viruses, including multiple coronaviruses like
SARSCOVE two, alpha and beta, human coronavirus H one N one,

(46:12):
and influenza B, influenza viruses hepatitis C, as well as
neurovirus and this battery of viruses gave us a significant
breadth of viral families to evaluate efficacy. And what we
found is indeed the drug did work across those disparate
viral families, but that still didn't mean that that we

(46:34):
had a drug. So what we had to do is
move from being AI enabled drug developers to really being
traditional drug developers. And from that point we took the
drug into in vivo models of efficacy with our collaborators
at Mount Sinai, and what we found is that the

(46:56):
drug indeed inhibited the symptom of SARS cove two in
a beautiful dose dependent manner. In this case, the animals
in question lose about twenty percent of their body weight
and it's a symptomatic sign of disease, and we were
able to alleviate that. And then of course, we went

(47:17):
to the primary biomarker, how much virus is existing in
the lungs of these animals, and we were able to
reduce the viral load in the lungs of these animals
by two point seven logs. Now, for those familiar with
the space, there was the Pfizer team published a paper
in Science early on the pandemic demonstrating that pex lovid

(47:41):
reduced viral tiders in the lungs of mice and a
equivalent animal model to ours to the tune of one
point three to one point nine logs. So with that,
we were very excited about the potency of our therapeutic,
but nonetheless we still didn't have a drug right so
there we needed to understand how much drug were we

(48:03):
delivering to the lungs, was it safe and well tolerated,
what the therapeutic indexes were, so on and so forth.
And since then we've done quite a bit of that
work and all of it has come back as positive.
And so right now in this initial program, we appear
to have a broad spectrum therapeutic with activity across a

(48:27):
significant swath of RNA viral biology. We appear to have
pre clinical proof of concept in vivo in a gold
standard animal model, and all of these results are able
to be explained with healthy animal PK studies, lung partitioning coefficients,

(48:50):
plasma protein binding studies, et cetera, et cetera. So, and
now this program is now marching to the clinic, and
we're really excited to bring this to the market. As
a doctor Davey Smith, one of our clinical advisors in
the head of Clinical Trials and Research that UCSD puts it.
You know, he actively still sees infectious disease patients and

(49:14):
there's not a therapeutic when somebody enters the emergency department
or his office with obviously viral respiratory symptoms that he
can start dosing them immediately in the same way there
are these therapeutics in the anti bacterial space, but they
don't exist in the anti viral space, and he's cautiously

(49:36):
hoping that this delivers a tool to clinicians that allows
them that same anti viral tool as we have in
the antibacterial space.

Speaker 1 (49:47):
So may just a few clarifying questions here. You describe
the thumb one region of RNA depend on RNA polimeraates
as a novel target, but it was targeted by Bristol Myers.
But klabevir alba for appetitis cevirus, So what makes you
view it as novel?

Speaker 2 (50:04):
Yeah, so so this is a this is a great question.
So but klavivier, as you mentioned, developed by BMS, approved
approved in Japan. Uh, Like like many drugs early on
in the pandemic, was tested for activity against stars Cove two,

(50:27):
and in fact, a number of physics based AI driven
models predicted that mclavivier would indeed work UH for SARS
Cove two, but that was found not to be in
fact the case UH in multiple UH in multiple studies. UH. Famously,

(50:49):
you know, Johnson Johnson did a high through but screen
of a number of studies and found it to be
inactive UH, as well as a number of other laboratory
library based high through PLUT screening studies, and so there
was there was no crumb to follow as far as
that was concerned. Also, if you look into the literature,

(51:14):
it had been hypothesized that the thumb one pocket existed
across viral biology potentially, but never had been determined to
definitively be the case. And the details get quite detailed
if you will, But the reason it was discovered on

(51:35):
hepatitis C is that, in fact, within viral biology RNA
viral biology, it's really the easiest VIRL RNA viral polymerase
UH to work with and to understand, and there are
some experimental limitations due to the detailed biology of polymerases

(51:57):
associated with other viruses that made it very difficult to
confirm that this pocket existed. And then when the lead
drug that had been developed for hepatitis C someone RDRP
failed in these other viruses, I think a lot of
scientists determined there was no reason to go down that route.

(52:19):
What's important for us, and this sort of gets back
to my idea that the AI frees us from dogma
is that is that when you have this plethora of
data stacked up against you, and you're a human scientist
working in a lab and there's only so many hours
of the day, you would never typically go down that
pathway and say, you know, my gut tells me that

(52:42):
this is going to work. I believe others have missed something,
so I'm going to keep proceeding. People tend to stop
at that point. However, when you're executing on AI driven
drug discovery, we are not necessarily limited in those same
ways that human scientists are limited. So we're able to say,

(53:05):
I know that this is what the data says to
this point, but I believe that there are reasons that
the data says that, and I don't believe that we
have a full data set relative to this idea in
front of us, and I'm going to go and test

(53:28):
this in the cloud in my platform, and it only
cost me. I mean, there's still real cost, but it
only cost me hours or days of time. And in
that case, that's what we did right. We didn't accept
that the data that was in the literature was the
full data set. We assumed that there were reasons for

(53:52):
that result, and it allowed us to still move forward
with this as an idea, and we were found to
be right. So, in fact, if you look at one
of our preprint papers and you evaluate, our model correctly
predicts that baiclavivier will work and hepatitis C and will
work well. And it also correctly predicts that the clavivial

(54:15):
will not work for SARS cove two, which is correct
unlike MDL zero zero one, which it predicts will be
active for both at therapeutic levels. Right, And this gets
back to this idea I was mentioning earlier in the podcast,
talking about the ability. It's one thing to discover something

(54:36):
that works. It's another thing to be able to definitively
look at a compound and say that is not going
to work. Right, and when you know but both what
will work and what will not work simultaneously, you control
that biochemical space, and when you can control that biochemical

(54:59):
s you can take risk that others cannot. On top
of being able to take risk others cannot because your
laboratory is the cloud and you can run thousands and
millions of experiments in a matter of hours to days
to weeks at most. And so so I think that
is That's something that I would love for the listeners

(55:21):
to understand, is that because of our approach, we're able
to not accept. We don't have to accept data at
face value. We can assume that there's more there and
we can continue to push, which allows you to discover

(55:42):
things that humans given the same data set or same
context wouldn't be able to get to.

Speaker 1 (55:51):
That makes a lot of sense. Then we know MDLS
erar zero one at this point is pipe and doxafine,
so it has activity and the gen receptors. So how
do you think about the potential for diminished anti viral
efficacy relative to your assay work due to on target
binding and humans to estrogen receptors since trying to bind
to both at the same time.

Speaker 2 (56:10):
I guess sure, sure, So again this is you know,
this is a great point and this is what this
this topic is really what AI drug discovery is built
on so or what it's for, right and what it
can do. So from one point of view to you know,

(56:32):
want to be clear for all the all the listeners
is that MDL zerser one pippendoxapen has seen clinical trials
phase one to be exact, it's been in two well
controlled phase one trials and it's been in almost one
hundred healthy patients without a report of grade three or

(56:55):
grade four adverse events above and beyond placebo. Right, So,
I do want to be clear that even in the
clinic pipendoxa fene demonstrated no issues related to safety and
tolerability in the healthy population that it was tested in.

(57:16):
I also want to say we Model Medicines have put
this drug in over one hundred and fifty animals at
doses up to one thousand mixx per kig, which is
a substantial dose without any adverse events observed. Right that

(57:37):
being said, just because all that's true, still doesn't mean
that we at Model Medicines think MDL zero zero one
and our infectious disease program is a finished product. We
believe that we can use our tools even if MDL
zero zero one is a significant drug that significantly either

(58:05):
treats or prevents the next pandemic or the various endemics
that affect human beings on a yearly basis, we still
believe we can do better. And one of the things
that we believe that we can always do better, whether
it's serm activity or any other specific drug activity, is

(58:25):
that we want to create drugs that are specific as
possible to the target that they are intended to drug. Right,
So put aside any details of what nonspecific activity that
they have, just period point blank, we want to make
the most drugs that are the most specific to the
target of interest. Right. So in this case, what we've done,

(58:48):
again going back to this idea of being able to
engineer both the positive space and the negative space are
very important to us. So in this case, what we've done.
We recently published the paper where not only did we
discover new chemical entities that bind to the thumb one

(59:09):
target and are active against multiple viral infections, but we
created compounds that specifically lacked simultaneously serm activity and so
and we ran those models at the same time to
reach those discoveries. So let me back up and say
what again. What we did. We modeled, if you will,

(59:33):
the positive chemical space of thumb one because we wanted
to discover drugs active against that target, and we also
modeled the negative space of the estrogen receptor, so that simultaneously,
what the model nominated were compounds that both had anti

(59:56):
viral activity and lacked sermactivity at the same time. And
the proof is in the putting here. When we synthesized
those compounds and took them into the lab, all twelve
of them had anti viral activity at therapeutic concentrations, and
all twelve of them had substantially reduced serm activity. And

(01:00:21):
to give you a sense of the level of reduction
of serm activity lead the smallest reduction was just shy
of eight hundred and thirty x reduction in estrogen receptor activity,
up to nearly sixteen thousand x reduction in estrogen receptor

(01:00:44):
activity simultaneously. The implication of this is that as we
move forward as a company and as drug developers, we're
not just going to be focused on eliminating estrogen receptor activity.
In this case, We're going to only be looking for
chemistry that eliminates all non specific interactions that in any

(01:01:10):
way could be detrimental to this drug working in as
perfected as in a way as possible, and then you
can even extrapolate from there that if you can do that,
you can also really really dial in individual interactions that allow
a drug to get into one tissue and not another preferentially,

(01:01:32):
or any other variable that you would like to model.

Speaker 1 (01:01:35):
And that's the future the thumb one program and the
model platform, or generally, let's talk about the future of
mblzer or zero one for a second. Give us an
update on your entry to clinics such as next steps
or tru design plans. You're going to be a focus
on potentially opportunistic infections and at risk individuals, or you're
going to try to go more broadly. I don't think
there's any real regulatory pathway for a universal anti viral.

Speaker 2 (01:02:00):
Yeah, so so this is this is also a very
interesting question and one that is something that is being
hotly debated inside of of model medicines. And there's there's
some really interesting pathways that you could go down in

(01:02:25):
terms of clinical programs that we're exploring with our clinical
advisory advisory team. And I'll just give you some of
the options that are there. So some interesting ideas, for
instance include, so you could do straightforward right, we have

(01:02:47):
the animal model work in sars coov two, we have
this two point seven log reduction in viral load, and
so you could do a clinical trial directly there. You
might also say, you know, just last winter, UH, we
had the triple demic it some had coined it where

(01:03:09):
we had these overlapping three overlapping respiratory viral infections UH
that were hitting young and old alike, leading to co
infections that were filling our emergency departments. And so a basket,
a sort of basket respiratory trial could also be in

(01:03:34):
in the mix here. The other interesting thing, too is
with the hepatitis C. With the hepatitis C H activity,
is that you know, it's not a great situation if
somebody has chronic hepatitis C and then is dealing with
a viral respiratory infection simultaneously, and there may be actually

(01:04:00):
what might be interesting is first getting clinical approval of
a co infection scenario that might be considered a rare
worth in indication, but one of which is really important
of course, and this is where the pharmaceutical industry overlaps
with real life. Of course, there are significant human migrations

(01:04:24):
that are occurring right now, and of course we'll stay
out of the politics of why this is happening, but
nonetheless it is and when we as drug developers or
clinicians simply have to deal with it. And these migration
patterns have led to a number of co infection scenarios
that we haven't really dealt with before, and so maybe

(01:04:47):
one of the more interesting ways things that we should
do where we can have the greatest impact, is to
go this route. And so I mean this all to
say that there are a number of different clinical development routes,
potentially rare and orphan co infection models going simply directly
for influenza or sarskov two as it might be there

(01:05:11):
in front of us, and we're currently debating all of those.
I think the biggest thing that I'd like for any
listener to take away from this though, is that there
are there are many reasons why you might do one
clinical trials versus another, but ultimately we exist as model medicines.

(01:05:35):
We are here every day using this platform to discover
drugs to eliminate human suffering, and ultimately what we want
to do is that when approved as safe, well tolerated,
and efficacious in a fully vetted clinical trial program, we

(01:05:58):
want these therapeutics to get as many people as possible
to solve human suffering. But how we get from here
to there? There are a number of variables at play.
There's a number of ideas, and I think, uh, experts
in the field might disagree, you know, reasonable minds could
disagree what's the best way, But ultimately our goal is

(01:06:22):
to deliver safe, well tolerated, and effective drugs to as
many people as possible, as soon as possible. Uh, And
that's really what our what our guiding principle is here.

Speaker 1 (01:06:33):
And maybe can we just close out with two general questions,
where do you think we're currently at on the Gardner
hype cycle with regards to AI and drug development?

Speaker 2 (01:06:43):
So we we are definitely we are definitely going through
a cycle right now, and and I think we are
we are certainly in the phase where uh, there is
uh some players in this space are starting to be
washed out, frankly, and you do see a number of

(01:07:08):
companies that are closing or emerging, and you do see
an elevated, elevated critiques of the industry that's occurring. There
have been a number of articles recently written by some
of the best journalists covering this space calling into question

(01:07:34):
whether the Emperor has any clothes in this space, And frankly,
we welcome these critiques, and we welcome them because we
believe that this industry of AI enabled drug discovery should
not be evaluated on area under the curved scores or

(01:08:01):
terabytes of data or supercomputer specifications. We believe that this
industry should be evaluated in time appropriately on how many

(01:08:21):
drugs have achieved pre clinical proof of concept, how many
drugs have achieved I in d how many drugs have
achieved phase one, Phase two, phase three, and eventually NDA clearance.
And I think what's going on now is that healthy
filtering of the system such that companies that have real

(01:08:47):
data because they have engineered thoughtful platforms, are going to
come to the four and platforms that have been less
successful in the creation of actual therapeutics will inevitably fade
into the background. So I don't think that's a bad thing.

(01:09:08):
I think that's a good thing. And we are definitely
in the part where the general media is critiquing where
companies are closing, but it's typically in this period where
those that will form the sustainable generation of companies fueled
by these technologies, they will come out of the wash,

(01:09:29):
and therefore they will also gain the proper support from
the venture capital community, from a big pharma and from
other supporters regulatory industries, etc. Etc. Because it's simply easier
to deal and work with a smaller set of companies
that have demonstrated their ability to achieve.

Speaker 1 (01:09:52):
Then what are you most excited about what's coming next?
Both for the model platform and the field more generally.

Speaker 2 (01:10:00):
Yeah, So, so I'll speak for for model very specifically.
So we've taken everything that we learned in the infectious
disease program that I alluded to earlier in the podcast,
and we brought that to our oncology platform and just
to give you, just to give you a sense of
what of what we've done. So in our original models,

(01:10:24):
we were discovering drugs like MDL zero zero one and
the NCE library that I referred to before, and we
were doing that with data sets that were measured, you know, say,
up to ten thousand. In our oncology program right now
where we're marching down the pathway, but potentially the we've

(01:10:49):
improved our platform such that we have discovered data in
the in the public realm that's ten x the size
of the data that we use to generate our ID program.
And so what's what's really exciting to me is an

(01:11:10):
oncology program to begin with. But what's really exciting to
me is I know we're even going to have better
results faster because we've gotten better at what we do,
and we literally see it in the ten x scene
of our data pipeline for instance. I'm also frankly really
excited in the industry generally as as drugs that have

(01:11:36):
entered more drugs entered the clinical development space, they're obtaining
I n ds and moving through the clinical development process.
Of course, some drugs now have solidly reached phase two
in this space that we're truly AI generated, not AI assisted,

(01:11:57):
in their discovery. And I look forward to very soon
people won't question whether AI can drive the discovery of
a drug that is solving disease for real patients. The
question will not be if it can occur. It will
be how many times can we repeat it, how many

(01:12:18):
companies can't repeat it, and how short of a development
timeline can we have and I think we're just ahead
of that precipice about to flip over, because there's some
really great entrepreneurs and really really great scientists in this
space leading really great companies as well. And I'll just
maybe end on this too. The pharmaceutical industry never has

(01:12:42):
and never will be, you know, have one company to
rule them also to speak, even though we are very
committed to our approach for AI driven drug discovery, it
inevitably will not be the only successful approach, and we
look forward to a future where multiple discipline, serious AI

(01:13:07):
driven drug discovery companies are transforming human health simultaneously. Because
also as any of us know who are part of
you know who are part of families, which we all are,
inevitably there's plenty of disease to go around, there's plenty
of human suffering to solve. And I look forward to

(01:13:28):
that discipline, serious group of companies taking their place in
pharmaceutical development and super rather than later alleviating that real
human suffering.

Speaker 1 (01:13:40):
Great, well, we'll be watching for updates from both your
virology and oncology pipelines. Thank you, Daniel for joining me
for this episode of Vanguards of Healthcare.

Speaker 2 (01:13:50):
Thank you for having me Andrew really really appreciate your
questions and really appreciate your expertise in this space, with
an honor to join you today

Speaker 1 (01:14:02):
Absolutly and thank you everyone for listening.