Latest / Vanguards of Healthcare by Bloomberg Intelligence / Xaira’s War on Trial-and-Error Drug Development

Transcript

Xaira’s War on Trial-and-Error Drug Development · Vanguards of Healthcare by Bloomberg Intelligence

0:26
Speaker 1: Welcome to another episode of Bloomberg Intelligence Vangers of Healthcare podcast,
0:30
Speaker 1: where we speak with the leaders at the forefront of
0:33
Speaker 1: change in the healthcare industry. My name is Sam Fazelli
0:35
Speaker 1: and I'm a healthcare analyst at Bloomberg Intelligence, the in
0:38
Speaker 1: house research arm of Bloomberg.
0:41
Speaker 2: I am thrilled.
0:42
Speaker 1: I think I say this every time, but I'm really
0:44
Speaker 1: thrilled to welcome Marcus Levigne, the.
0:47
Speaker 2: CEO of Zera, to the program today.
0:49
Speaker 1: Mark has put me right in how to pronounce that
0:52
Speaker 1: the spelling is x ai r A so Zera. Mark
0:56
Speaker 1: has more than decade well, he has decade of experienced
1:00
Speaker 1: as a scientist in the world of drug development, running companies,
1:04
Speaker 1: founding companies, and here we find him leading this AI
1:09
Speaker 1: drug Development drug discovery company. Welcome to the podcast.
1:12
Speaker 3: Mark, Thank you, Sam, it's great to be here.
1:15
Speaker 2: Thank you.
1:16
Speaker 1: I've been at the in the audience where Mark has
1:18
Speaker 1: been presenting in the scientific days, so I'm really pleased
1:21
Speaker 1: with that. Now, before we get going, I always like
1:25
Speaker 1: to get a little history of the of the folks
1:28
Speaker 1: that we talk to, and no less I think Mark,
1:31
Speaker 1: you did say your PhDs was at actually in my hometown,
1:34
Speaker 1: which is London, and we obviously crossed paths at Guy's Hospital.
1:38
Speaker 1: But do you want to give us a little few
1:40
Speaker 1: minutes of background as to how what the process that
1:43
Speaker 1: got you to here is today?
1:45
Speaker 2: Great?
1:45
Speaker 3: Well, thank you Sam again, it's it's great to be here.
1:48
Speaker 3: Thank you for having me. I'm really excited to be
1:51
Speaker 3: leading Zara now focused on applying AI to transform how
1:54
Speaker 3: we discover and develop drugs. I came to it through
1:58
Speaker 3: several chapters. I started my career as a scientist, as
2:02
Speaker 3: you mentioned. I came my PhD in neurophysiology at University
2:06
Speaker 3: College London, then moved to the US for the first time.
2:09
Speaker 3: I'm Canadian, so it was my first move to the
2:12
Speaker 3: US to do postdoctoral work in neuroscience and took my
2:15
Speaker 3: first job at UCSF in San Francisco, and there my
2:19
Speaker 3: focus was entirely on how the brain gets wired up
2:23
Speaker 3: during embryonic development. Like you, at the time, I was
2:26
Speaker 3: passionate about understanding how nerve cells connect with other nerve
2:30
Speaker 3: cells in the brain, and we and others in the
2:33
Speaker 3: field were able to make a lot of progress and
2:35
Speaker 3: identifying brain wiring molecules. Molecules like netans for example, to
2:39
Speaker 3: be be discovered. But as we discovered fundamental mechanisms, for
2:43
Speaker 3: a number of reasons, some personal my father had a stroke.
2:48
Speaker 3: I became interested in applying those discoveries to try to
2:52
Speaker 3: rewire the nervous system after injury, and that led me
2:56
Speaker 3: into more of a translational route. Thing led to another.
3:01
Speaker 3: I realized that after co founding a startup company in
3:05
Speaker 3: back in two thousand with a colleague in the field,
3:08
Speaker 3: that I would have a bigger impact if I moved
3:11
Speaker 3: to the private sector myself and was recruited to gen
3:14
Speaker 3: and Tech. That was really the start of my work
3:16
Speaker 3: in drug discovery. I joined genen Tech, overseeing about two
3:21
Speaker 3: thirds of the research enterprise, eventually running the entire research
3:24
Speaker 3: and drug discovery enterprise, and I was very fortunate to
3:28
Speaker 3: learn from the best of the best in the industry there.
3:31
Speaker 3: This was in the two thousands. Eight FD approved drugs
3:33
Speaker 3: came from projects that I helped oversee, and I would
3:38
Speaker 3: have continued there except I was recruited back to wonderful
3:41
Speaker 3: positions in leadership in academia at Rockfeller University and then Stanford University.
3:46
Speaker 3: Even while I kept a hand in biotech, being on
3:50
Speaker 3: the boards of some companies co founding another company focused
3:54
Speaker 3: on nerdegeneration, an area of obvious passion for me from
3:58
Speaker 3: my scientific days, and all of this led me to
4:01
Speaker 3: this point where, when I was still at Stanford, I
4:04
Speaker 3: was approached by the three founders of Zera. The company
4:09
Speaker 3: is really focused on applying AI to drug discovery. This
4:12
Speaker 3: is the moment when the advances in AI and biological
4:16
Speaker 3: insights have come together to really, we believe, make it
4:20
Speaker 3: possible to revolutionize how it is that we discover and
4:25
Speaker 3: develop medicines. And I just became very excited about this
4:29
Speaker 3: opportunity to bring my experience in biological sciences and drug
4:33
Speaker 3: discovery together with leaders in AI to embark on this
4:38
Speaker 3: journey to pioneer the use of AI to fundamentally transform
4:42
Speaker 3: how we discover and develop new medicines. So that is
4:46
Speaker 3: what motivated me to join these three founders. Of the
4:49
Speaker 3: three founders, I should say a Nobel Laureate, David Baker,
4:52
Speaker 3: Bob Nelson from Arch Ventures, and Vic Bajaje and Fortside Capital,
4:55
Speaker 3: they come together to conceive of a company that would
4:58
Speaker 3: apply AI to drug discovery and to d and approached
5:00
Speaker 3: me about being the founding CEO. So I just feel
5:04
Speaker 3: very privileged at this moment in time when the advances
5:07
Speaker 3: in AI have are transforming every domain to be able
5:11
Speaker 3: to bring that together with my background in biology and
5:15
Speaker 3: drug discovery to try to make a difference for patients
5:18
Speaker 3: we're suffering from poorly treated diseases.
5:21
Speaker 1: So that's fascinating to a mark and I think it
5:24
Speaker 1: gives you the insight and background to the drug the
5:30
Speaker 1: history of the drug development world that should be able
5:33
Speaker 1: to help me, certainly and our listeners to understand the
5:39
Speaker 1: way that AI is today changing things and potentially can
5:43
Speaker 1: change things forward for drug discovery and drug development. But
5:46
Speaker 1: before we do that, I think a natural question is
5:49
Speaker 1: David Baker won the Notebook, shared the Nobel Prize with
5:52
Speaker 1: Emisa Saves for the protein AI use in protein folding prediction.
6:00
Speaker 2: Would you want to give.
6:02
Speaker 1: A slight give us some idea of how perhaps ZERA
6:07
Speaker 1: differs in terms of approach to what you know and
6:11
Speaker 1: publicly is knowledge is known about isomorphic approach there, because
6:15
Speaker 1: I guess the background knowledge that fed both companies is
6:20
Speaker 1: relatively similar.
6:21
Speaker 3: Yeah, well, let me start by maybe just mentioning again
6:25
Speaker 3: or building on what you just said about the Nobel
6:28
Speaker 3: Prize in the fall of twenty twenty four, the Nobel
6:33
Speaker 3: Prize was shared by two groups, Demosis Sabasin, John Jumper
6:37
Speaker 3: at Google deep Mind, and they got half the price
6:41
Speaker 3: for Alpha fold for showing how you can use AI
6:45
Speaker 3: to predict how proteins. For proteins are as you know,
6:49
Speaker 3: made of strings of building blocks amino acids. It's been
6:53
Speaker 3: believed for close to seventy years that just the sequence
6:57
Speaker 3: of the proteins would determine how the proteins fold. With
7:01
Speaker 3: matters for drug discoveries the shape of proteins, because that's
7:04
Speaker 3: how they do their work. That's also how we make drugs.
7:07
Speaker 3: They attached to particular parts of a protein and the
7:09
Speaker 3: shape matters enormously for that so being able to predict
7:13
Speaker 3: shape from sequence was very important, and they showed how
7:17
Speaker 3: you could use AI leveraging a very large data set.
7:21
Speaker 3: This is important actually to where the field is called
7:24
Speaker 3: the Protein Data Bank. It's two hundred thousand highly curated
7:30
Speaker 3: structures of proteins that have been assembled over a period
7:34
Speaker 3: of about seventy years, many by academics, some by industry,
7:38
Speaker 3: deposited in a public repository for everybody to use. An
7:43
Speaker 3: alpha fol trained on that data set and incredibly showed
7:48
Speaker 3: that you could use that information with AI to predict
7:52
Speaker 3: with high accuracy the structure of most proteins, and for
7:58
Speaker 3: that they received half of the the Nobel Prize. The
8:01
Speaker 3: other half went to David Baker, or co founder. Building
8:06
Speaker 3: on that work and also on work that he himself
8:09
Speaker 3: had done previously, he showed how you could use generative
8:12
Speaker 3: AI to create entirely new proteins of desired functions. So
8:17
Speaker 3: it's a synthetic biology if you will, to say, can
8:21
Speaker 3: we use AI to design a protein that has a
8:22
Speaker 3: particular shape and a particular function, for example, to attach
8:26
Speaker 3: itself to a drug target in order to switch it
8:29
Speaker 3: off or to switch it on again, one of the
8:32
Speaker 3: key steps in making drugs. So David showed how one
8:36
Speaker 3: could use AI to do that and they shared the
8:38
Speaker 3: Nobel Prize. So that gives you the backdrop. There's this
8:42
Speaker 3: very large database that can be trained on to predict
8:45
Speaker 3: the shape of proteins and how proteins interact. It can
8:49
Speaker 3: be used also with further refinements, to create entirely new proteins,
8:54
Speaker 3: and that is what has enabled this explosion of activity
8:58
Speaker 3: around applying AI to a particular aspect of drug discovery.
9:03
Speaker 3: So here, Sam Fit's okay, I'm going to walk you
9:05
Speaker 3: through the drug discovery more generally and where this piece fits,
9:10
Speaker 3: and then I'll come back to your question about how
9:13
Speaker 3: what we're doing, you know, differs from what isomorphic and
9:17
Speaker 3: others are doing.
9:18
Speaker 2: That'd be perfect Mark.
9:20
Speaker 1: And also in the process, could you also touch on
9:22
Speaker 1: because there are some other companies, I mean, one of them,
9:24
Speaker 1: Generate Biomedicines, is pitching for an IPO at the minute,
9:29
Speaker 1: just contrast the approaches perhaps, or maybe maybe it's too
9:34
Speaker 1: much for a podcast today, but I'd love to hear
9:37
Speaker 1: how you think everything fits together with regards to the
9:42
Speaker 1: current scene of protein design AI driven protein design.
9:47
Speaker 3: So that's great, Let's take it in stages, Sam Fit's, okay.
9:51
Speaker 3: With you. First, let's look at the drug development process
9:54
Speaker 3: as a whole. You have to do three things to
9:57
Speaker 3: make a drug. First, you have to choose the molecular target.
10:01
Speaker 3: Typically it's a protein in the body that has a
10:04
Speaker 3: particular function and it might be a good actor, it
10:07
Speaker 3: might be a bad actor, and you're going to make
10:09
Speaker 3: a drug that will have an intended effect. So, for example,
10:12
Speaker 3: the GLP one drugs that everybody's familiar with like ozepic,
10:17
Speaker 3: they function to switch on a good actor, a protein
10:21
Speaker 3: called the JLP one receptor in the cells in the
10:25
Speaker 3: body cells, so that's switching on a good actor. If
10:29
Speaker 3: in cancer often there's a bad actor, a protein that's
10:32
Speaker 3: been gone awry through mutation, where what you want with
10:35
Speaker 3: the drug is to switch it off. So you first
10:38
Speaker 3: have to say what's the molecular target. Then you have
10:40
Speaker 3: to make a drug that has the intended effect typically
10:42
Speaker 3: of switching it on or switching it off. And then
10:45
Speaker 3: a third key step is to actually match the drug
10:49
Speaker 3: and the target to the right patients. You can have
10:51
Speaker 3: the best drug in the world against the best target.
10:53
Speaker 3: If you go into clinical trials and give it to
10:57
Speaker 3: the wrong patients who aren't going to benefit, then you
11:00
Speaker 3: will fail. In fact, a lot of the failure in
11:01
Speaker 3: drug discovery occurs at that stage where you might have
11:04
Speaker 3: a pretty good drug against a good target, but you
11:06
Speaker 3: don't know who the patients are are responding. So those
11:09
Speaker 3: three steps choosing the right target, making the right drug,
11:12
Speaker 3: and matching to the right patients. Currently, our industry really
11:17
Speaker 3: approaches us as an artisanal process. A lot of science,
11:22
Speaker 3: of course at each stage, but a lot of empiricism,
11:24
Speaker 3: trial and error. A lot of art people talk about
11:27
Speaker 3: intuition in making the drugs, and as a result, there's
11:32
Speaker 3: very high failure rates, very long timelines, and very great expense. Currently,
11:36
Speaker 3: to go from choosing molecular target to having an FDA
11:39
Speaker 3: drug approved drug on average takes about twelve years, and
11:44
Speaker 3: ninety percent of the drugs entriclinical trials fail and that
11:48
Speaker 3: is a broken process that we have to improve at
11:51
Speaker 3: the highest level. The promise of AI is that with
11:54
Speaker 3: the right kinds of data in the right amounts, we
11:58
Speaker 3: should be able to transform each of those three steps
12:01
Speaker 3: so that much more of the work can be done
12:02
Speaker 3: in silico on the computer, with higher success rates, lower attrition,
12:07
Speaker 3: shorter timelines, lower expense. So that's the vision for the
12:12
Speaker 3: industry as a whole. So when we look at how
12:16
Speaker 3: AI is being applied to this process and what the
12:18
Speaker 3: different companies are doing, we are actually focused on all
12:21
Speaker 3: three steps. I'll tell you how many companies are focused
12:24
Speaker 3: on one or other of those steps, and I haven't
12:28
Speaker 3: even mentioned, you know, a fourth application of AI. So
12:31
Speaker 3: there's AI applied to choosing targets, AI applied to making drugs,
12:35
Speaker 3: AI applied to patient stratification, identifying the right patients to
12:39
Speaker 3: treat with the drug. There's a fourth step which I
12:42
Speaker 3: won't talk about because we're not focused on that ourselves,
12:44
Speaker 3: which is just using AI to improve logistics and operations
12:48
Speaker 3: choosing clinical trial sites for example, or helping you know,
12:52
Speaker 3: investigators write applications to the FDA or reports things like that.
12:58
Speaker 3: That's happening in a very rapid way. People have adopted
13:03
Speaker 3: the tools, large language models and so forth to accelerate
13:06
Speaker 3: and improve operations and logistics, just as they are in
13:09
Speaker 3: every single industry, not just healthcare. So that's terrific. That
13:13
Speaker 3: will drive productivity gains, and there's some great companies that
13:16
Speaker 3: are doing great work there. That's not what we're focused on.
13:18
Speaker 3: We're focused on those three steps, the scientific steps of
13:21
Speaker 3: target drug, patient identifying the right ones. So will AI
13:27
Speaker 3: help in all three of those For choosing molecular targets.
13:32
Speaker 3: What we need to do is develop biology models, foundation
13:36
Speaker 3: models of biology that understand biology at a causal level. Currently,
13:41
Speaker 3: if you have a cell in a disease state and
13:43
Speaker 3: you want to bring it back to a healthy state,
13:46
Speaker 3: the way we go about trying to figure out what
13:48
Speaker 3: molecular changes to make to do that are typically involve
13:51
Speaker 3: a lot of trial and error, sometimes some kinds of
13:54
Speaker 3: molecular screening, but they're limited in impact, and we believe,
14:00
Speaker 3: I think, with a lot of justification as as the field,
14:02
Speaker 3: that again, with the right kinds of data, by generating
14:05
Speaker 3: a lot of causal data, we should be able to
14:07
Speaker 3: set up a causal model of cells. So we can
14:09
Speaker 3: address that question in silicon on the computer what set
14:13
Speaker 3: of changes to make to bring the cell from a
14:16
Speaker 3: disease state to a healthy state, which should improve target identification.
14:20
Speaker 3: In terms of making drugs, there are actually some quite
14:23
Speaker 3: good methods for making drugs currently, typically that involve high
14:27
Speaker 3: throughput screening of various kinds. I think your listeners will
14:30
Speaker 3: have heard of high throughput screening. You have a protein,
14:32
Speaker 3: you want to find a chemical that attaches to it,
14:34
Speaker 3: or an antibody at biologic so you screen large libraries
14:38
Speaker 3: of compounds or large libraries of biologics, often in the
14:41
Speaker 3: form of phage or other types of displays, and you're
14:46
Speaker 3: looking for the needle in the haystack, the one that
14:49
Speaker 3: will attach itself to the drug target to have the
14:51
Speaker 3: intended effect. The promise of AI is that with generative AI.
14:55
Speaker 3: Rather than looking for the needle in the haystack, we
14:57
Speaker 3: can design the needle. Much as you would prompt you know,
15:01
Speaker 3: Sora to make a movie for you. You enter a
15:03
Speaker 3: prompt about you know you want the AI to make
15:06
Speaker 3: a movie of two puppies playing on the beach. Here
15:09
Speaker 3: you can say you can prompt the AIS that we're
15:11
Speaker 3: generating and others are generating, to say, I'd like to
15:14
Speaker 3: make an antibody a biologic to you that will attach itself,
15:18
Speaker 3: that will bind to the target protein in just the
15:20
Speaker 3: right way to have the intended effect. So that's the
15:23
Speaker 3: second step applying AI there. The third step of applying
15:27
Speaker 3: AI to matching drugs to patients. Identifying who which patients
15:31
Speaker 3: are most likely to respond involves other kinds of data.
15:35
Speaker 3: Data on longitudinal data sets that follow the progression of
15:40
Speaker 3: disease and patients match to various kinds of phenotypic markers.
15:45
Speaker 3: And the promise there is that with the right kinds
15:48
Speaker 3: of data on the right amounts, we should be able
15:49
Speaker 3: to create patient representation models, where again we can ask
15:53
Speaker 3: in silico which patients are the ones most likely to
15:56
Speaker 3: respond to the drug to help guide clinical trial develop
16:00
Speaker 3: meant to increase the success rate there. So that's where
16:04
Speaker 3: AI is being applied. In those three areas, different companies
16:09
Speaker 3: are focusing on different aspects of the process. In the
16:12
Speaker 3: case of Isomorphic and our Cells, where a lot of
16:17
Speaker 3: the focus there is on the second one, making drugs.
16:19
Speaker 3: We've come from different areas. Isomorphic started working on a
16:23
Speaker 3: certain class of therapeutics. As you know, they are different
16:27
Speaker 3: therapeutic modalities. There are small molecules pills, if you will,
16:30
Speaker 3: There are biologics, antibodies. There are other kinds of modalities
16:36
Speaker 3: gene therapy, cell therapy, and so forth. We've been focusing
16:41
Speaker 3: our Cells on large molecule therapeutics, on antibody therapeutics, because
16:45
Speaker 3: that's where we have made the most progress in terms
16:48
Speaker 3: of the developing the technology. Isomorphics started with small molecules.
16:53
Speaker 3: Most companies that are working on one modality are interested
16:56
Speaker 3: in working in other modalities as well. That's true of us,
16:58
Speaker 3: it's true of them. So currently we're not immediately overlapping,
17:02
Speaker 3: but I expect there's going to be a convergence. And
17:05
Speaker 3: what you have is a number of companies trying to
17:08
Speaker 3: develop algorithms to do the best job possible of developing
17:12
Speaker 3: drugs in silico against molecular targets. If I can take
17:17
Speaker 3: suggest a few more words about what we're doing at
17:19
Speaker 3: ZERA and sort of the business proposition. When it comes
17:23
Speaker 3: to using generative AI to make drugs to make antibodies,
17:28
Speaker 3: there are a few things you could do. One is
17:30
Speaker 3: to use the model to try to make drugs faster
17:34
Speaker 3: against targets that you could against which you could develop
17:37
Speaker 3: drugs by other means. So there currently are some very
17:41
Speaker 3: good methods for making drugs and they work very well
17:44
Speaker 3: against certain classes of proteins. And again, if you indulge me,
17:50
Speaker 3: maybe it can explain to listeners. Think of a protein
17:53
Speaker 3: that looks sort of like a stick that's sticking out
17:57
Speaker 3: of the cell, and what you want to do is
18:00
Speaker 3: a drug that will attach itself to the part of
18:03
Speaker 3: the protein that's sticking out of the cell. It's relatively
18:06
Speaker 3: straightforward to make that portion of the protein and screen
18:10
Speaker 3: it in one of these high throughput screening approaches, and
18:14
Speaker 3: you don't necessarily need AI to get a drug. What
18:18
Speaker 3: we believe over time is that AI should enable us
18:20
Speaker 3: to get some of those drugs faster, but we're not
18:22
Speaker 3: there yet. But what you can do with AI that
18:25
Speaker 3: you can't easily do with current methods is to go
18:27
Speaker 3: after proteins that are more difficult to study. Think now
18:31
Speaker 3: of a cell membrane and think of a protein that
18:33
Speaker 3: snakes in and out of it the membrane, where only
18:36
Speaker 3: the tip of the iceberg, the tip of the protein
18:38
Speaker 3: is sticking out of the cell membrane. It's very difficult
18:41
Speaker 3: to screen for that kind of protein for drugs that
18:43
Speaker 3: will attach them cells in that way. AI is not
18:46
Speaker 3: limited in that way because you ask the AI to
18:49
Speaker 3: make a drug against whatever part of the protein you want,
18:52
Speaker 3: and so it's not limited by the difficulties of screening
18:55
Speaker 3: those proteins. So their entire classes of proteins that are
18:59
Speaker 3: well known to be good drug targets. Some of them
19:02
Speaker 3: have names like GPCRs g protein coupled receptors. They snake
19:05
Speaker 3: across the membrane multiple times. They've been a great set
19:09
Speaker 3: of proteins to go after for drug development, but many
19:13
Speaker 3: of them are difficult to screen and certainly difficult to
19:16
Speaker 3: screen for antibody therapeutics. We're focusing on those kinds of targets,
19:21
Speaker 3: targets that have some or a lot of validation where
19:24
Speaker 3: we believe we could make a real difference to patients
19:26
Speaker 3: if we only we could make a drug against it.
19:28
Speaker 3: But that are difficult to screen with existing methods. So
19:33
Speaker 3: if you will, we're not trying to go for the
19:34
Speaker 3: low hanging fruit faster. We're trying to go for the
19:37
Speaker 3: high hanging fruit that's difficult to access with current methods.
19:40
Speaker 3: So AI we believe will enable us to do two things.
19:43
Speaker 3: One is to go after difficult to drug targets, the
19:46
Speaker 3: so called undruggable targets. That's where we're focused. But also,
19:50
Speaker 3: over time, as our models get better and better, to
19:52
Speaker 3: just make drugs faster than you can by conventional methods.
19:56
Speaker 3: So both of those are the promise.
19:58
Speaker 2: For that was brilliant, Mark, thank you very much. Let
20:03
Speaker 2: me let me pick through this a little bit.
20:05
Speaker 1: We have the cell, the protein target, and the data
20:08
Speaker 1: mm hmm, right to.
20:10
Speaker 2: Sell the cell.
20:11
Speaker 1: Now I've read Andrew White has made a estimate that
20:15
Speaker 1: you need ten to the thirty eight flops to simulate
20:19
Speaker 1: a human cell. Do you need do you think to
20:22
Speaker 1: simulate the every single atom to be able to make
20:27
Speaker 1: a in silico cell useful or can you do it?
20:32
Speaker 1: Or do you not need to go to that level?
20:34
Speaker 2: Of perfection.
20:35
Speaker 3: I think you've framed it perfectly, Sam, that what we
20:38
Speaker 3: need to do is have a model of the cell
20:41
Speaker 3: that is sufficiently rich that you can actually get meaningful
20:44
Speaker 3: and useful answers. But it doesn't necessarily have to capture
20:47
Speaker 3: every single detail on the cell, at least at this
20:50
Speaker 3: point in time. And know as scientists our ambition will
20:54
Speaker 3: be ultimately to do that, but we don't have to
20:57
Speaker 3: reach out ultimate goal to be able to make meaningful
21:00
Speaker 3: progress in developing models that can be useful for drug discovery.
21:04
Speaker 3: So what do you need to make a model of
21:07
Speaker 3: cells that is useful for drug discovery? So let me
21:10
Speaker 3: tell you the approach that we're taking currently with There
21:13
Speaker 3: are a lot of data in the public domain, generated by academics,
21:17
Speaker 3: generated by industry that are descriptive data of cells. People
21:22
Speaker 3: have taken cells of a particular kind. They might be
21:24
Speaker 3: liver cells, they might be muscle cells or nerve cells
21:29
Speaker 3: of interest, and they've identified within those cells every single
21:35
Speaker 3: one of the genes that is turned on. You know,
21:38
Speaker 3: we have about twenty thousand genes, give or take. Each
21:42
Speaker 3: gene codes for at least one protein, in some cases
21:47
Speaker 3: a few versions of a protein. So twenty thousand genes,
21:50
Speaker 3: about twenty thousand proteins, and there's an intermediate. When a
21:55
Speaker 3: gene is switched on to make the protein the cell,
21:58
Speaker 3: it goes through an intermediate called mRNA. You know, your
22:01
Speaker 3: listeners would not have been familiar with mRNA twenty years ago,
22:04
Speaker 3: but after COVID and the mRNA vaccines, everybody has heard
22:07
Speaker 3: of mRNA. So mRNA is the messenger between a gene
22:10
Speaker 3: which is you know, unchanged in every cell, but when
22:13
Speaker 3: it gets switched on, it will it will be read
22:16
Speaker 3: initially as mRNA. Then the mRNA gets converted to proteins,
22:20
Speaker 3: it gets read out to make proteins. The what you
22:24
Speaker 3: can do with cells is a very powerful technology for
22:26
Speaker 3: identifying all the mRNAs in the cell and which tells
22:29
Speaker 3: you all the genes that are switched on and by inference,
22:32
Speaker 3: all the proteins that are being made, which are more
22:34
Speaker 3: difficult to quantify and measure. So you can take cells,
22:38
Speaker 3: measure all the mRNAs and you can ask of the
22:41
Speaker 3: twenty thousand genes in the cell, how many are switched on.
22:45
Speaker 3: People have done that experiment in many different cell types
22:48
Speaker 3: and cell types that are healthy that are diseased. Put
22:50
Speaker 3: them in the public domain, and what you end up
22:51
Speaker 3: with is very descriptive data of the state of cells
22:56
Speaker 3: at particular times. Using AI, it's possible to make models
23:00
Speaker 3: of cells that are descriptive. In fact, our head of
23:03
Speaker 3: biomedical AI Bo Wang, created the first such model right
23:07
Speaker 3: out on the heels of chat GPT. He created single
23:10
Speaker 3: cell GPT SCGPT, which is an AI model that relies
23:15
Speaker 3: on these public domain descriptive data to create a model
23:19
Speaker 3: that's very good at descriptive tasks for cell type annotation,
23:23
Speaker 3: identifying batch effects, or other things of a descriptive nature.
23:26
Speaker 3: But if you ask the causal question that I started
23:28
Speaker 3: with a few minutes ago, if the cell is at
23:31
Speaker 3: a disease state, how do you bring it to a
23:34
Speaker 3: healthy state? What changes do you have to make in
23:36
Speaker 3: the cell to do that? What causal actions are necessary
23:40
Speaker 3: for that? Those models fail completely. And the reason they
23:43
Speaker 3: fail is because they're trained on descriptive data. So you
23:47
Speaker 3: train on descriptive data. You're good at descriptive questions, you
23:50
Speaker 3: fail at causal questions. What's needed are causal data. There
23:55
Speaker 3: are various ways of creating causal data. One of the
23:57
Speaker 3: ways we're doing it is using a technology that enables
24:00
Speaker 3: you to modify in cells one gene at a time,
24:04
Speaker 3: so switching off genes. So you take a cell, you
24:06
Speaker 3: might take, as we've done, a liver cell, or immune cell,
24:10
Speaker 3: or a cancer cell, and switch off each of the
24:14
Speaker 3: twenty thousand genes individually in those cells and read out
24:17
Speaker 3: what's the effect on all twenty thousand genes. So it's
24:20
Speaker 3: a twenty thousand gene perturbation experiment by twenty thousand gene readout.
24:25
Speaker 3: Those are the kind of data, causal data that are
24:27
Speaker 3: necessary to create a model that starts to understand that
24:30
Speaker 3: the genetic networks in the cell at a causal level.
24:34
Speaker 3: So we have published and we are generating massive amounts
24:39
Speaker 3: of causal data to train causal models of the cell
24:44
Speaker 3: to be able to ask that question, what does it
24:46
Speaker 3: take to go from a health a disease state to
24:49
Speaker 3: a healthy state. And we believe that those kinds of
24:52
Speaker 3: data will be sufficient in the right quantities and the
24:55
Speaker 3: right cell types will be sufficient to be able to
24:58
Speaker 3: address questions that are useful therapeutically. Again, just to answer
25:02
Speaker 3: your question, do we need to have a complete model
25:05
Speaker 3: the cell to be able to make progress? The answer
25:09
Speaker 3: is no. We believe that this will be a first
25:13
Speaker 3: draft version of the cell. If you will first draft
25:15
Speaker 3: causal model on which we're going to model layer much
25:19
Speaker 3: more data, for example, not just which mRNAs are there
25:22
Speaker 3: and how they interact, but which proteins are actually present
25:27
Speaker 3: and other features of the cell. So I think as
25:29
Speaker 3: a whole, the field as a whole, we'll be able
25:32
Speaker 3: to make progress by creating appropriate causal models. It will
25:35
Speaker 3: become more and more and more accurate, richer and richer
25:39
Speaker 3: in data to provide better and better answers. But we
25:43
Speaker 3: believe that in the near term will be able to
25:44
Speaker 3: get some answers from these initial causal models.
25:47
Speaker 2: Of the cell. That's great, thank you.
25:48
Speaker 1: I was hoping you would say that, because I don't
25:50
Speaker 1: think we have enough energy to run something that requires
25:53
Speaker 1: into the thirty eight flops of computation. Yeah.
25:56
Speaker 3: Absolutely, But so.
25:58
Speaker 1: Mark on that specific point, A lot of diseases caused
26:01
Speaker 1: not I mean, the perturbation may be on a single
26:04
Speaker 1: cell level, but there's a network effect that that comes
26:08
Speaker 1: in in terms of an organ basis or aracrime, a
26:11
Speaker 1: variety of ways that that that the biology is more
26:16
Speaker 1: complicated than a single cell level. Where do you think
26:19
Speaker 1: we are actually two questions here. Where do you think
26:21
Speaker 1: we are in terms of those what you just spoke
26:25
Speaker 1: about in terms of having those models functionally useful already,
26:28
Speaker 1: and where do you think we are in terms of
26:30
Speaker 1: trying to get the network to be able to pertable
26:34
Speaker 1: look at that.
26:35
Speaker 3: That's a great question and exactly the kind of path
26:37
Speaker 3: that we're on. So the story of AI is you
26:42
Speaker 3: train on massive amounts of data that you can I
26:46
Speaker 3: was going to said that are easily available. None of
26:47
Speaker 3: us is easily available, but there's some that you can
26:50
Speaker 3: garner in very large, large amounts in relatively straightforward manner.
26:54
Speaker 3: As we have been with these perturbation experiments, Those experiments
26:58
Speaker 3: done in cells, and especially experiments done in cells in
27:01
Speaker 3: petri dish, are removed from the ultimate goal of understanding
27:07
Speaker 3: what's happening in a human being, and we have various
27:10
Speaker 3: models of increasing degrees of complexity. You go from single
27:13
Speaker 3: cells to organoids, which of course is another technological advance
27:17
Speaker 3: in the past decade, or even into animal models where
27:22
Speaker 3: you can do experiments perturbation experiments in the living organism
27:26
Speaker 3: in a mouse, for example, and finally to the human,
27:30
Speaker 3: where it may not be possible to do that directly
27:32
Speaker 3: in the human but you may be able to use surrogates,
27:35
Speaker 3: certainly use tissues from patients. With AI, what you do
27:40
Speaker 3: is you first build a model with massive amounts of
27:42
Speaker 3: data that are readily available. Then you can that's the
27:45
Speaker 3: pre training. You can then post train with data that
27:49
Speaker 3: might be closer to what you're most interested in that
27:51
Speaker 3: are harder to come by, and you can do further
27:53
Speaker 3: fine tuning with data that are very hard to come by,
27:58
Speaker 3: like disease tissues from from patients. And the extraordinary thing
28:03
Speaker 3: about AI is to see the transfer learning that occurs
28:06
Speaker 3: between those different levels. Let me give an analogy with chatchebt.
28:12
Speaker 3: Chatchipt trained on all the world's information, but in its
28:20
Speaker 3: prior to post training and fine tuning, it was not
28:23
Speaker 3: really very useful as a chatbot. It had to be
28:26
Speaker 3: taught how to use the information that was there in
28:29
Speaker 3: order to answer the queries that you might have. And
28:32
Speaker 3: we know that you can train further on other data
28:34
Speaker 3: sets that are even more highly curated and more accurate,
28:39
Speaker 3: for example, for legal documents or other things, to create
28:43
Speaker 3: various kinds of AIS that are more adept at certain
28:46
Speaker 3: types of task than just chatchept itself. In the same way,
28:50
Speaker 3: we can create models of the cells and then add
28:53
Speaker 3: information through this post training and fine tuning using tissues
28:58
Speaker 3: and data that are hard to come by, and that
29:01
Speaker 3: is how I believe we'll make progress to get at
29:04
Speaker 3: what you asked about, which is, how do you look
29:07
Speaker 3: not just at cells in a Patriot dish, which is
29:09
Speaker 3: very rarefied, but at cells in their natural habitat where
29:12
Speaker 3: they're interacting with other cells, which will also influence the
29:15
Speaker 3: course of the disease.
29:17
Speaker 1: So that's again we've talked about cell We've talked about networks,
29:20
Speaker 1: We've talked about protein targets on the proteins in terms
29:23
Speaker 1: of developing.
29:24
Speaker 4: Perhaps either being able to use the modified proteins as
29:29
Speaker 4: therapeutics themselves, but specifically in that case, do you how
29:33
Speaker 4: do you avoid the problem of immunegenicity, because that is
29:38
Speaker 4: the further you get away from the what I would
29:41
Speaker 4: call natural biology try and modify something, the higher the
29:45
Speaker 4: risk that you'll have that Or do you not believe that?
29:49
Speaker 3: No, I absolutely believe that, and this is an issue.
29:52
Speaker 3: There are various kinds of ways of making biologic therapeutics.
29:56
Speaker 3: Antibodies the you know you do it in and just
30:00
Speaker 3: a strait mouse, you'll end up with immune genesty problems
30:03
Speaker 3: because the mouse antibodies aren't quite the same as human antibodies,
30:06
Speaker 3: and the human immune system will detect those differences and
30:11
Speaker 3: will mount an immune response that will reject the therapeutic.
30:15
Speaker 3: This is a problem that had to be overcome when
30:17
Speaker 3: antibodies were first developed as therapeutics over thirty years ago.
30:21
Speaker 3: So what we've learned from that is that you can
30:23
Speaker 3: learn some of the rules of immunogenousity. That people make
30:26
Speaker 3: modifications of mouse antibodies, you can bypass the issue altogether.
30:30
Speaker 3: A decade later, people made mice where they replace the
30:32
Speaker 3: mouse's immune system with a human immune system, and those
30:36
Speaker 3: give you fully human antibodies. That's one way to go,
30:39
Speaker 3: but there are other ways as well. In making people
30:41
Speaker 3: have experience making synthetic antibodies and then humanizing them, so
30:46
Speaker 3: that is certainly one route. There are various assays. There's
30:50
Speaker 3: been a big effort by the industry as a whole
30:52
Speaker 3: over the past decade to develop assays that enable you
30:56
Speaker 3: to determine whether a protein is immunity. Some of it
31:01
Speaker 3: involves in silicoanalysis. Some of it involves, for example, looking
31:05
Speaker 3: at uptake of the protein that you've made, the synthetic
31:08
Speaker 3: protein by the sentinel cells of the immune system, the
31:11
Speaker 3: dendritic cells. There are other you know, in vitro assays
31:15
Speaker 3: like that that you can use to determine whether your
31:17
Speaker 3: protein is immunogenic or not, and then you can modify
31:20
Speaker 3: it to fix it. The beauty of AI, of course,
31:23
Speaker 3: is that you can not just modify to fix if
31:26
Speaker 3: you the antibody that you've made in the first instance
31:29
Speaker 3: is immunogenic. You can also generate massive amounts of data
31:33
Speaker 3: on synthetic antibodies to determine which ones are most likely
31:37
Speaker 3: to be immunogenic and which ones are less likely to
31:40
Speaker 3: be immunogenic, and then bake that into your models so
31:43
Speaker 3: that from the get go you create antibodies that are
31:45
Speaker 3: less likely to be immunogenic. So there are a number
31:48
Speaker 3: of properties of antibodies that you want to make a
31:51
Speaker 3: good antibody, and one that might be obvious is you
31:55
Speaker 3: don't want the antibody to aggregate because it might just
31:57
Speaker 3: you know, fall out of solution, and that's something that
32:01
Speaker 3: you can teach with appropriate data. So some of the
32:03
Speaker 3: physical chemical properties of antibodies, we are generating data to
32:07
Speaker 3: teach our models how to create antibodies that don't have
32:10
Speaker 3: those liabilities. I think the hardest one is the ones
32:13
Speaker 3: you identified, which is immunogenicity, because we don't have a
32:16
Speaker 3: perfect way of reading that out for physical chemical properties.
32:19
Speaker 3: We can read it out in the lab, so it
32:21
Speaker 3: can generate massive amounts of data and have a high
32:23
Speaker 3: degree of assurance that will be able to overcome that
32:25
Speaker 3: problem altogether using AI for immunogenicity. It's going to be
32:29
Speaker 3: semi empirical initially, but we think we can make a
32:31
Speaker 3: lot of progress and again a combination of AI driven
32:36
Speaker 3: modifications to the proteins and perhaps still some art of
32:40
Speaker 3: looking at the sequences and saying, well, you know, maybe
32:42
Speaker 3: this is more likely to be immunogenically with the kind
32:46
Speaker 3: of implicit knowledge that people have developed over the past
32:48
Speaker 3: several decades, is what will happen in the first instance.
32:51
Speaker 3: Over time, I think we will overcome that too, but
32:54
Speaker 3: you're right that that is probably one of the biggest
32:57
Speaker 3: areas that we'll have to attend to.
33:00
Speaker 1: So I'm going to make an assertion here or a
33:02
Speaker 1: statement for you to tear apart AI. Application of AI
33:08
Speaker 1: is going to be more fruitful in the hands of
33:11
Speaker 1: a large farmer company than a biotech Why large farmer
33:16
Speaker 1: companies have at least in certain therapeutics areas where they've
33:21
Speaker 1: been gathering data and collecting information, generating data, lab notes, etc.
33:27
Speaker 2: For decades.
33:29
Speaker 5: Assuming that data is going to be available to interrogation
33:35
Speaker 5: by AII, it's been correctly annotated, it's sort in the
33:39
Speaker 5: right place, they should have a better chance of using
33:43
Speaker 5: AI in terms of training AI for their models than
33:48
Speaker 5: a smaller startup or buyota company, etc. That's the statement.
33:52
Speaker 5: What's your reaction.
33:54
Speaker 3: I think there will be lots of opportunities for both
33:57
Speaker 3: big farmers and small companies. And for the small companies,
34:01
Speaker 3: I do think that there are opportunities to lead in
34:04
Speaker 3: development of the technology, but also great opportunities to partner
34:08
Speaker 3: with pharma partners to bring together the technology from the
34:11
Speaker 3: small company with both the expertise and the as you
34:16
Speaker 3: point out, the data that the large pharma might have.
34:20
Speaker 3: So I see this more as an opportunity for synergies
34:24
Speaker 3: rather than a you know, sort of winner take all
34:27
Speaker 3: kind of situation, which is the way you've phrased the question.
34:31
Speaker 3: We're certainly very mindful of that. You know, I think
34:33
Speaker 3: all of us as startups don't want to, you know,
34:36
Speaker 3: invest our resources in activities that you know, we're just
34:42
Speaker 3: replicating what others have done previously. We're trying to create,
34:46
Speaker 3: for example, novel data sets for foundation models of biology.
34:51
Speaker 3: By the way, those kinds of data sets don't exist
34:53
Speaker 3: in any pharma right now, just to be clear, So
34:56
Speaker 3: for the target identification side of the equation, the data
35:02
Speaker 3: pretty much has to be generated. What pharma can bring
35:05
Speaker 3: to the table there often is deep insights into certain
35:08
Speaker 3: therapeutic areas, great models, a lot of expertise and experience
35:13
Speaker 3: in studying those models, both in rarefied settings, in vitro
35:18
Speaker 3: in a Petritish but also in the living organism, and
35:21
Speaker 3: to match them with the kind of technology that we're
35:24
Speaker 3: applying in our studies. So I think there's an opportunity
35:27
Speaker 3: for synergy there, but I don't think this is a
35:30
Speaker 3: case of pharma having all the answers. I agree with you.
35:33
Speaker 3: On the clinical trial side, many pharmers have a terrific
35:37
Speaker 3: experience in clinical trials and have a lot of data.
35:41
Speaker 3: Often the data is not in the right format. I
35:45
Speaker 3: may be possible to convert it for use, but there
35:48
Speaker 3: are two things about the data. Also, in the clinical
35:50
Speaker 3: setting that are worth pointing out. The first is that
35:52
Speaker 3: the data that are going to be most valuable again
35:56
Speaker 3: will be prospective data where you have made a decision
36:00
Speaker 3: in advance to test patients in one setting or you know,
36:04
Speaker 3: for example, with one drug or not having a drug,
36:07
Speaker 3: and then measuring deeply the response in the relevant tissues.
36:12
Speaker 3: Sometimes that's been done in the pharmacy. Sometimes that has
36:15
Speaker 3: not that they don't have the biospecimens from the patients
36:19
Speaker 3: in order to feed to feed into the AI models
36:23
Speaker 3: to try to identify who's likely to respond and who's
36:25
Speaker 3: not likely not to respond. So there may be data
36:29
Speaker 3: that are useful, there are probably a lot of data
36:31
Speaker 3: that are not as useful as it might appear at
36:33
Speaker 3: base value, and there's going to be a need for
36:36
Speaker 3: additional data that's generated again in very tightly controlled clinical
36:41
Speaker 3: settings to address specific issues that can then be fed
36:45
Speaker 3: to the models to get the answers that we need
36:49
Speaker 3: who's most likely to respond. So I think there's going
36:52
Speaker 3: to be a combination of all of that in terms
36:54
Speaker 3: of you know, the drug making per se. There's clearly
36:58
Speaker 3: some pharmers have data where they've looked at the structure
37:02
Speaker 3: of small molecule drugs together with proteins, those are going
37:06
Speaker 3: to be very valuable for further training data sets. So
37:10
Speaker 3: there is data within the industry that can be accessed,
37:14
Speaker 3: but a lot of data will have to be generated.
37:17
Speaker 3: A lot of it will be bringing together certain technologies
37:20
Speaker 3: and certain AI models that are being developed in startups
37:24
Speaker 3: to bear on those data sets in collaborative actions. I
37:27
Speaker 3: think you're going to see a whole range of a situation.
37:30
Speaker 3: Sam You'll see some big pharmas that might be able
37:33
Speaker 3: to go it alone. There are many that will look
37:35
Speaker 3: to partner with smaller companies that have advanced the technology
37:39
Speaker 3: to a point where they want to take advantage of that,
37:41
Speaker 3: and the small company looks forward to being able to
37:45
Speaker 3: work with the data sets that the partner has.
37:48
Speaker 1: Okay, that was a perfect answer to something that keeps
37:50
Speaker 1: going around my head, and I keep asserting that point
37:52
Speaker 1: and I'll stand corrected. Talking about data, one of the
37:56
Speaker 1: problems for training models on public data is that there
38:01
Speaker 1: is a risk that negative data experiments that didn't work
38:06
Speaker 1: are not published. Now, this is not my great idea.
38:09
Speaker 1: I think a lot of people have.
38:11
Speaker 2: Raised this issue.
38:12
Speaker 1: How do you deal with that when you're training in
38:15
Speaker 1: AI on say everything has ever been published on PubMed
38:19
Speaker 1: or pubcap or whatever.
38:22
Speaker 3: I think it's a real issue, and I think that's why,
38:25
Speaker 3: just to tie back to your previous question, why some
38:28
Speaker 3: of the data that pharma companies or biotech companies or
38:31
Speaker 3: others have in their vaults may not be as useful
38:35
Speaker 3: as data that are obtained in prospective trials. It's absolutely
38:39
Speaker 3: the case that the AI will learn best if it's
38:42
Speaker 3: given both the positive results and the negative results, and
38:46
Speaker 3: that way you can reinforce the model towards the positive
38:50
Speaker 3: results away from the negative results. That is why many
38:55
Speaker 3: of us are generating data even for for example, antibody
38:58
Speaker 3: protein interaction. There are a number of antibody structures that
39:03
Speaker 3: are present in the database. Is a very small fraction
39:06
Speaker 3: actually of the two hundred thousand highly curated protein structures.
39:10
Speaker 3: But those are obviously the ones that were successful, and
39:12
Speaker 3: we haven't seen all the ones that were not successful,
39:15
Speaker 3: and so we are so that has some value, absolutely,
39:19
Speaker 3: but much more values derived will be derived by generating
39:23
Speaker 3: data sets where you have thousands upon thousands of antibody
39:28
Speaker 3: structures and ideally closely related and you see which ones work,
39:33
Speaker 3: which ones combined with high affinity and which ones don't.
39:37
Speaker 3: That's the kind of data that the model will learn
39:39
Speaker 3: best from. So your point is correct that there's a
39:43
Speaker 3: paucity of negative data out there, but we can definitely
39:46
Speaker 3: make advances with just positive data. Alpha fold was trained
39:51
Speaker 3: on the protein data bank and there was enough richness
39:55
Speaker 3: of interactions there and I should say cleverness of the
40:00
Speaker 3: Mind team that they were able to extract the information
40:03
Speaker 3: or train the AI to predict the structures of most proteins.
40:07
Speaker 3: So it is one can certainly make advances based just
40:10
Speaker 3: on positive data, but there's no question that one can
40:13
Speaker 3: make further advances and perhaps more rapid advances, by having
40:18
Speaker 3: data both the positive and the negative data.
40:20
Speaker 1: Great, thank you, so Mark, I'm want to ask you
40:23
Speaker 1: very quickly to answer a question as simply as you can,
40:26
Speaker 1: and then move on to the drug areas or therapeutic
40:30
Speaker 1: areas where you think zero would be focused on going forward,
40:33
Speaker 1: and then also tell.
40:34
Speaker 2: Me where you would hope you would be when you
40:38
Speaker 2: would be able.
40:39
Speaker 1: To file an I in D or how many you expected,
40:41
Speaker 1: So a whole bunch of questions in terms of future looking.
40:45
Speaker 1: First question is though your twelve years approval target to approval.
40:49
Speaker 1: How much can we shave off that? Just an estimate,
40:52
Speaker 1: and then talk to me about please, the therapeutic areas
40:56
Speaker 1: that Zara is going to be focused on, if you've
40:58
Speaker 1: decided already, and also the when when you think that
41:02
Speaker 1: you might be in a position to file at least
41:04
Speaker 1: the first I ND and then of course everybody says
41:07
Speaker 1: we're going to file an ID three IDs a year,
41:09
Speaker 1: you know, those sort of pontifications.
41:11
Speaker 2: I'd love to hear if you want to make that.
41:13
Speaker 3: Yeah, well, thank you, Well let's start. And I'm glad
41:16
Speaker 3: that you asked the question the way you did about
41:19
Speaker 3: you know, the twelve years of making drugs, because I
41:23
Speaker 3: think what we should do is set ourselves a goal
41:26
Speaker 3: that is highly aspirational but also realistic, something that we
41:31
Speaker 3: could achieve. I would like to see us And currently
41:35
Speaker 3: the data it takes about twelve years actually thirteen years
41:39
Speaker 3: on average to make a drug. That's five years to
41:42
Speaker 3: go from target to making a drug that enters clinical trials,
41:47
Speaker 3: and then on average eight years to conduct the clinical
41:50
Speaker 3: trials and get FDA approval. So total of thirteen years
41:53
Speaker 3: and ninety percent and depending on the statistics that you
41:57
Speaker 3: look at, ninety to ninety five percent of the drugs
42:00
Speaker 3: enter clinical trials will fail and not make it over
42:02
Speaker 3: the finish line, which let's put in terms of the
42:05
Speaker 3: success rate there it's you know, let's say ten percent
42:07
Speaker 3: of the drug set inenter clinical trials will make it
42:09
Speaker 3: over the finish line. I think we should set a
42:12
Speaker 3: goal for ourselves of by over the next decade, reducing
42:16
Speaker 3: the time from targetifty approval in half six and a
42:20
Speaker 3: half years, and multiply the success rate from ten percent
42:24
Speaker 3: on average to something realistic that would make a huge
42:27
Speaker 3: difference would be thirty percent, thirty to forty percent. I'd
42:30
Speaker 3: love it to be fifty percent, but even a thirty
42:32
Speaker 3: percent success rate would make a huge difference. It would
42:36
Speaker 3: It would change the rate of approvals and also dramatically
42:40
Speaker 3: reduce the costs the economics of drug discovery. You know,
42:43
Speaker 3: this has been analyzed in detail, changed dramatically if you
42:46
Speaker 3: can improve the success rates by that kind of amount.
42:49
Speaker 3: So I would say cut in half and increase success rate,
42:53
Speaker 3: you know, triple the success rate would be a huge
42:56
Speaker 3: step forward, and I think is not unrealistic if you
43:00
Speaker 3: look at each of the steps. So let's just take,
43:02
Speaker 3: for example, the first step of making the drug. So
43:05
Speaker 3: from choosing a target to making a drug and getting
43:08
Speaker 3: into clinical trials what you currently five years. We should
43:13
Speaker 3: be cutting that in half or even less if we
43:17
Speaker 3: can essentially make medicines in silico much more rapidly to
43:21
Speaker 3: go from target to a drug candidate in a year
43:25
Speaker 3: to a year and a half. There's always going to
43:27
Speaker 3: be some trial and error and testing and then being
43:30
Speaker 3: able to get into the clinic in a short period
43:34
Speaker 3: of time. Again, that's something that people have been working
43:37
Speaker 3: on currently. You know, it takes over a year with
43:40
Speaker 3: an antibody on average to get into clinic, although for
43:43
Speaker 3: industry best is the below a year, and we there
43:47
Speaker 3: are enough examples of keeping people going fast, including during
43:50
Speaker 3: COVID when some medicines were developed very rapidly, that we
43:54
Speaker 3: can benchmark ourselves against that and say how can we
43:56
Speaker 3: do that systematically, So it's not just a one off
43:58
Speaker 3: in terms of the clinic trials. Reducing the time it
44:02
Speaker 3: takes in clinical trials is all about patients stratification. If
44:07
Speaker 3: you can better identify the patients who are going to respond,
44:11
Speaker 3: your trial doesn't have to be as big. You know,
44:13
Speaker 3: if in a thousand patients they're only two hundred that
44:18
Speaker 3: are going to respond. You know that may not be
44:21
Speaker 3: enough to give you statistical significance. If you already know that,
44:24
Speaker 3: you know you can identify patients who have more than
44:26
Speaker 3: a fifty percent chance of responding. You don't have to
44:29
Speaker 3: treat as many in the clinic. You'll see fewer patients enrolled,
44:33
Speaker 3: larger effect sizes, and obviously higher success rates and shorter timelines.
44:38
Speaker 3: So I think that's what we should be aspiring to
44:41
Speaker 3: in terms of our own therapeutic areas and timelines. And
44:46
Speaker 3: this gives me the opportunity to say one way in
44:47
Speaker 3: which we are distinct from many companies out there, not
44:51
Speaker 3: all of them, is that we aren't just developing the technology.
44:54
Speaker 3: Many small companies are developing the technology and then seeking
44:58
Speaker 3: to partner or let since that technology. We're developing the technology.
45:02
Speaker 3: We will do that as well, but we're also making
45:05
Speaker 3: our own therapeutics, and as I mentioned, we're focused on
45:09
Speaker 3: large molecule therapeutics that are currently undruggable or hard to
45:14
Speaker 3: drug by conventional methods. We have not described our therapeutic
45:18
Speaker 3: areas of interest or our timelines. So since we haven't
45:22
Speaker 3: done that, I'm not going to do that here today, Sam.
45:24
Speaker 3: But but I can tell you is this first, that
45:26
Speaker 3: we are focused on those hard to drug targets, the
45:29
Speaker 3: high hanging fruit. As a biotech company, it's important to differentiate.
45:33
Speaker 3: If you work on targets that are of high value
45:36
Speaker 3: but that are straightforward to drug by conventional means, there'll
45:39
Speaker 3: be a dozen companies working on those targets and it's
45:42
Speaker 3: hard to differentiate. So we're going after the hard targets
45:45
Speaker 3: where people have not been successful at making drugs, and
45:49
Speaker 3: that's what's driving our choice of targets and therapeutic areas currently.
45:52
Speaker 3: Targets with some are a lot of validation that have
45:55
Speaker 3: been hard to drug by conventional means, and so we
45:58
Speaker 3: are in multiple therapeutic areas currently, and we are pushing
46:03
Speaker 3: very hard to get to a clinic as rapidly as possible.
46:07
Speaker 3: But that's all I'm willing to say about timelines at
46:09
Speaker 3: this point.
46:09
Speaker 2: Okay, Thanks Mark.
46:10
Speaker 1: I mean, look, we've been talking for coming up to
46:13
Speaker 1: about an hour, so that's very generous of you, and
46:15
Speaker 1: I know I was a bit late.
46:17
Speaker 2: I apologize again.
46:18
Speaker 1: Do you want to say some final words here? I
46:20
Speaker 1: think of I mean, I can keep talking to you
46:22
Speaker 1: about this subject and keep drilling and drilling and drilling.
46:25
Speaker 1: But let me just give you the opportunity to wrap
46:28
Speaker 1: this up. And you've been involved in drug development for
46:30
Speaker 1: years and I just want to know you know, I mean,
46:33
Speaker 1: and you've got other companies you're on the board and
46:36
Speaker 1: shareholders and other companies you've founded, other companies. If you
46:39
Speaker 1: wanted to also end with telling me what's your genuine
46:43
Speaker 1: excitement about this? I mean, obviously you have some phenomenal
46:46
Speaker 1: backers Bob Nelson and the Foresight team and the arch
46:51
Speaker 1: Ventures team, so I mean you're in great company with
46:54
Speaker 1: also David Baker. What is the I mean, how excited
46:57
Speaker 1: are you about AI in general? So let's not talk
47:00
Speaker 1: about zero specifically delivering on its promise, because that is
47:04
Speaker 1: where I think technology sometimes has the risk of promising
47:09
Speaker 1: a lot and not always obviously delivering well.
47:12
Speaker 3: The you know, thank you for this opportunity. The for me,
47:16
Speaker 3: what I'm passionate about is, you know, applying science to
47:20
Speaker 3: develop medicines that can help patients who are in need.
47:25
Speaker 3: And it's been I've been very fortunate to be able
47:27
Speaker 3: to work with some of the best companies.
47:30
Speaker 1: You know.
47:30
Speaker 3: I was chief SCIENTIAG officer at Genentech. I was on
47:34
Speaker 3: the board of Regeneron for a number of years. I've
47:37
Speaker 3: had a front row seat to some of the best
47:42
Speaker 3: science and the best drug discovery in the industry for
47:46
Speaker 3: several decades now. Nothing excites me as much as the
47:51
Speaker 3: potential for AI, and that's why I just felt so
47:54
Speaker 3: fortunate to be approached about bringing my experience in drug
47:58
Speaker 3: discovery to the company. I really believe that this promise
48:03
Speaker 3: of AI, that that we've been discussing, that with the
48:07
Speaker 3: right kinds of data will be able to dramatically improve
48:11
Speaker 3: the success rates for all three steps of drug discovery
48:15
Speaker 3: is real. I believe it requires us to work extremely
48:19
Speaker 3: hard and make the right moves to generate the right
48:21
Speaker 3: kinds of data sets and apply them in the right way,
48:24
Speaker 3: not just to do proof of concept experiments. Oh, we
48:27
Speaker 3: can make something that looks like an antibody, that's great,
48:29
Speaker 3: But can you make an antibody that has all the
48:32
Speaker 3: properties of an antibody, including, as you pointed out, low immunogenicity.
48:36
Speaker 3: I believe the answer is yes. I think we can
48:39
Speaker 3: break down each of the barriers that's in our way,
48:42
Speaker 3: and that we should be able to do it in
48:45
Speaker 3: you know, with alacrity the we're so I couldn't be
48:49
Speaker 3: more excited myself. I couldn't be more excited about what
48:54
Speaker 3: where the field is going as a whole. And what
48:56
Speaker 3: we really need to do is to be smart and
48:59
Speaker 3: discipline and focus where it's going to make the biggest
49:02
Speaker 3: difference for each of those three steps of drug discovery.
49:05
Speaker 3: And the end goal, of course, is to be able
49:08
Speaker 3: to bring better medicines faster to patients and need. And
49:13
Speaker 3: I think AI is going to enable us to do that.
49:16
Speaker 3: If we continue the way we're going right now, it
49:18
Speaker 3: will continue to be trial and error. And as you know, Sam,
49:22
Speaker 3: the cost of making drugs, there's something called e Room's law,
49:26
Speaker 3: which is Moore's law and reverse which is instead of
49:30
Speaker 3: cutting the price of semiconductors every two years, it's doubling
49:35
Speaker 3: the price of drugs every two years because of continued
49:38
Speaker 3: inefficiencies and attrition. I believe we can stop that, and
49:43
Speaker 3: we can reverse that and have something closer to a
49:46
Speaker 3: Moore's law for medicines, and I think AI is going
49:49
Speaker 3: to play a major role in that. I should say
49:52
Speaker 3: it won't just be AI, will be the continued ingenuity
49:55
Speaker 3: of scientists doing thinking of new approaches as well new
49:59
Speaker 3: therapeutic fidalities and the like, and better ways of doing
50:03
Speaker 3: things across the board. But I'm very, very persuaded and
50:05
Speaker 3: confident that AI is going to play a central role
50:08
Speaker 3: in turning this important but somewhat broken approach and system
50:12
Speaker 3: that we have into something that's more of a finely
50:14
Speaker 3: tuned machine for delivering medicines for patients.
50:17
Speaker 1: That's wonderful, Mark, I'm really really enjoyed the whole conversation,
50:22
Speaker 1: but specifically the last thing I had not heard of
50:25
Speaker 1: e room law. But just to give you a feel
50:29
Speaker 1: for this, we're modeling. We were just talking to one
50:32
Speaker 1: of my colleagues the other day modeling a new drug
50:34
Speaker 1: and we were going back and forth about what tries
50:37
Speaker 1: to pick, and so I said, okay, let's just take
50:39
Speaker 1: the average of the past three years of oncology drug launcher.
50:43
Speaker 1: Is what price did they launch out on an annual
50:45
Speaker 1: cost spasis, leaving the car ties aside, because they're one, yes, Rey,
50:50
Speaker 1: well three hundred and fifteen thousand dollars, And you think
50:53
Speaker 1: to yourself, well, we can't keep going like this. I
50:56
Speaker 1: think Farmer needs to perhaps because their costs. But maybe
51:00
Speaker 1: if as you say, we can bring those costs down
51:03
Speaker 1: and the speed to market faster, then maybe they'd be
51:06
Speaker 1: incentivized to bring products to market.
51:09
Speaker 2: At a more at the lower price.
51:11
Speaker 1: I don't want to say it's more affordable cheaper, it's
51:13
Speaker 1: nothing to do with that, and also continue to add
51:16
Speaker 1: significant value.
51:17
Speaker 2: So that would be great to see.
51:19
Speaker 1: If Mark, I want to thank you again, we've I
51:22
Speaker 1: think we're definitely over the one hour mark and it's
51:24
Speaker 1: been a great opportunity to catch up with you and
51:28
Speaker 1: hear the story. I hope that in a year or
51:30
Speaker 1: two or even sooner, I mean AI moves very fast,
51:34
Speaker 1: we can get back together and talk about some progress
51:36
Speaker 1: that the company has made.
51:37
Speaker 3: Thank you, Sam, I'd love to do that, and I
51:39
Speaker 3: look forward to continuing the conversation.
51:41
Speaker 2: Wonderful. Thank you very much for your time, and look
51:44
Speaker 2: forward seeing you soon.
51:45
Speaker 1: Thank you. Uses as bases us
52:27
Speaker 3: Sissases