Latest / Vanguards of Healthcare by Bloomberg Intelligence / Inside Insilico Medicine’s AI Strategy and Lilly Breakthrough

Transcript

Inside Insilico Medicine’s AI Strategy and Lilly Breakthrough · Vanguards of Healthcare by Bloomberg Intelligence

0:27
Speaker 1: Welcome to another episode of Bloomberg Intelligence's Vanguards of Healthcare podcast,
0:31
Speaker 1: where we speak with the leaders at the forefront of
0:34
Speaker 1: change in the healthcare industry. My name is San Fazelli
0:37
Speaker 1: and I'm a pharmaceutical anainist of Bloomberg Intelligence, the in
0:40
Speaker 1: house research arm of Bloomberg. I am thrilled to welcome
0:46
Speaker 1: Alex Javoronkov, founder and CEO of In Silico Medicine. Alex,
0:51
Speaker 1: Welcome to the program.
0:53
Speaker 2: Happy to be with you, Sam.
0:54
Speaker 1: Today, as usual, I'm going to start asking Alex a
0:59
Speaker 1: little bit about himself. I had the privilege of spending
1:02
Speaker 1: some quality time together with Alex a couple of weeks
1:05
Speaker 1: back in Shanghai, and we had a good event there
1:09
Speaker 1: where you were a key contributor, which I'm really appreciative of.
1:13
Speaker 1: So Alex, tell us a little bit about yourself for
1:15
Speaker 1: our listeners, and you know how you started into the
1:19
Speaker 1: world of computing, computition, coding, whatever you like, and how
1:23
Speaker 1: it brought you to where you are today.
1:25
Speaker 2: Thank you. Sam. First of all, great to be on
1:28
Speaker 2: your podcast, and it was a truly enjoyable experience to
1:32
Speaker 2: be part of your event in Shanghai and deeply enjoy
1:36
Speaker 2: your research and also I can see how deep you
1:39
Speaker 2: guys go, right, So actually scheduling one on ones and
1:44
Speaker 2: group meetings with Kio wells in our respective industries in
1:48
Speaker 2: bio as well, and I actually learned a lot and
1:51
Speaker 2: met many new good people where I didn't know in
1:54
Speaker 2: China myself. So my name is Alex Chavran Kaff. I'm
1:58
Speaker 2: the founder and cosey CEO of in Silico Medicine, but
2:02
Speaker 2: for the past nine months I also were the head
2:05
Speaker 2: of a chief business officer off in Silico. So you've
2:09
Speaker 2: seen that we now have many more deals coming. So
2:12
Speaker 2: I really want to focus on ensuring financial sustainability for
2:17
Speaker 2: the company for many years to come. I am a
2:19
Speaker 2: computer scientist by training, so first two degrees from Queen's
2:23
Speaker 2: Unversity in Canada and computer science and commerce. Then worked
2:27
Speaker 2: for a number of years in semiconductor industry, actually in
2:30
Speaker 2: GPU computing, working for a company which used to compete
2:34
Speaker 2: with a Nvidia called ATI Technologies, and later it was
2:37
Speaker 2: acquired by AMD. I became more or less independently wealthy
2:41
Speaker 2: in my early twenties. I was one of the youngest
2:43
Speaker 2: directors at ATI and I was lucky enough to actually
2:47
Speaker 2: buy a bunch of fares in the right time and
2:50
Speaker 2: decided to commit the rest of my life to aging research.
2:54
Speaker 2: So I actually very much focused most of my activities
2:58
Speaker 2: kind of on target an extracurricular on the study of
3:02
Speaker 2: biological processes and time, which is basically studying aging, and
3:07
Speaker 2: did my grad work at just Hopkins, worked for a
3:10
Speaker 2: number of biotechs. I was a professor here and there,
3:13
Speaker 2: had a stint as an international adjunct faculty at the
3:17
Speaker 2: Moscow Institute Physics and Technology, among all places, and went
3:23
Speaker 2: back to Baltimore in twenty fourteen, established in Silico medicine
3:28
Speaker 2: and at the campus of the Emerging Technology Centers. At
3:32
Speaker 2: Hopkins ended around the time of Nvidio GtC, so I
3:36
Speaker 2: was deeply inspired by the image net competition and the
3:40
Speaker 2: early Atari Games paper by deep Mind and decided that, okay,
3:45
Speaker 2: we need to go and apply this technology deep learning
3:48
Speaker 2: that I was very familiar with at the time. I
3:50
Speaker 2: was very familiar with GPU computing because of my time
3:54
Speaker 2: at ATI, where I kind of cut the teeth in
3:57
Speaker 2: many areas, including GPGPU nowadays not a very used term,
4:02
Speaker 2: but at the time that's general graphics processing units, and
4:05
Speaker 2: I started working on scaling very early, and in two
4:09
Speaker 2: thousand and fifteen we published our first papers on utilizing
4:14
Speaker 2: deep learning for aging research for age prediction, and two
4:18
Speaker 2: thousand and sixteen published our first papers on the use
4:22
Speaker 2: of generative adversarial networks for small molecule drug design. Our
4:27
Speaker 2: first five years of the company existence, we were servicing pharma,
4:31
Speaker 2: trying to understand where in which kind of pockets deep
4:35
Speaker 2: learning can be effective and generative technologies could be effective.
4:39
Speaker 2: Twenty seventeen we synthesized our first small molecule generated by
4:46
Speaker 2: AI and synthesized and tested. Twenty eighteen, we went all
4:50
Speaker 2: the way into mice, resulting in a publication in Nature
4:53
Speaker 2: by Technology that made us pretty famous. Took almost a
4:57
Speaker 2: year in review, so we published in two thousand on
5:00
Speaker 2: nineteen and September the Technological Generative Tensorial Reinforcement Learning for
5:04
Speaker 2: Small Molecule drug Design. We're in twenty one days we
5:08
Speaker 2: generated six molecular structures for a given protein target, and
5:12
Speaker 2: in forty six days that includes that twenty one days
5:16
Speaker 2: we managed to get into mice with mouspikada completing binding assays,
5:21
Speaker 2: microsomal stability or metabolic stability assays. And went all the
5:25
Speaker 2: way into mice and that made us pretty famous. We also,
5:30
Speaker 2: for the first time at that time, encountered China. So
5:33
Speaker 2: I realized that, okay, well, where do we synthesize the
5:36
Speaker 2: chemistry for this experiment? And at that time there were
5:40
Speaker 2: several places to synthesize, but most of the two places
5:44
Speaker 2: where we could go were India and China where it
5:47
Speaker 2: could scale and do it cheaply with high quality, including
5:52
Speaker 2: binding assays, including anymore experiments, and we started collaborating with
5:55
Speaker 2: a company called Wushi Apptech. I was at that time
5:58
Speaker 2: very cash poor. I had always about six months of
6:01
Speaker 2: cash in a bank and most of the money came
6:04
Speaker 2: from me at some point in time because we kind
6:06
Speaker 2: of failed to attract capital externally and in the early days,
6:11
Speaker 2: and I had two rounds of synthesis left in the bank,
6:15
Speaker 2: so we had to go cheap, quick and fully integrated.
6:18
Speaker 2: So I went to China, met Wushi, saw how they
6:21
Speaker 2: do things. We agreed to do a pilot and kind
6:24
Speaker 2: of time the effort, and we realized that wow, we
6:27
Speaker 2: can actually produce reasonably good molecules and synthesize and tests.
6:32
Speaker 2: And that publication led to the thirty seven million dollars round,
6:36
Speaker 2: and we decided to actually pull a trigger on our
6:40
Speaker 2: own pipeline, so trying to go with our own drugs
6:44
Speaker 2: all the way to pre clinical experiments and also launch
6:48
Speaker 2: our software. So nowadays I referred to AI as a
6:53
Speaker 2: software for drug discovery, and we started selling our biology software,
6:58
Speaker 2: chemistry software and software for the prediction of clinical trial
7:02
Speaker 2: outcomes in clinical trial design in twenty twenty and also
7:05
Speaker 2: pulled the trigger on our own programs. So since then,
7:09
Speaker 2: Oh and another thing that happened in my life that
7:11
Speaker 2: changed it completely. I met doctor rerem my co CEO.
7:16
Speaker 2: Now at that time she joined as a CSO who
7:18
Speaker 2: had after Harvard eleven years at GSK and medicinal Chemistry
7:23
Speaker 2: in leadership roles, and then co piloted a CRO country
7:28
Speaker 2: research organization called Medicillin, which competed with WUSHI, and at
7:32
Speaker 2: that time it was high flying, a multi billion dollar company.
7:36
Speaker 2: CROs were on their peak, at their peak, and I
7:40
Speaker 2: managed to convince him to join in Silico to kind
7:43
Speaker 2: of turn AI into the Druggist discovery engine, because he
7:47
Speaker 2: really wanted to discover novel molecules and take them all
7:51
Speaker 2: the way to clinic. And when he was at GSK
7:53
Speaker 2: in eleven years. He nominated several pccs but couldn't take
7:57
Speaker 2: a single drug into phase one for many reasons, including
8:01
Speaker 2: strategic and here after joining in silicos Is twenty twenty one,
8:05
Speaker 2: when he joined officially, we nominated twenty eight developmental candidates
8:10
Speaker 2: and twelve of them reached clinical stage clearedynd and three
8:15
Speaker 2: Phase twos one phase to a complete and we published
8:19
Speaker 2: a bunch of papers together and one on them drug
8:22
Speaker 2: discovery and licensing partnerships. So that's what that's kind of
8:26
Speaker 2: my story in a nutshell. At some point in time
8:29
Speaker 2: in our history, we also spun off a company called
8:32
Speaker 2: Deep Longevity that focused and focuses on the development, design
8:37
Speaker 2: and development of aging clocks biomarkers of aging, and we
8:41
Speaker 2: sold this company to another listed company called Regent Pacific,
8:45
Speaker 2: which is now the owner of that company. They are
8:48
Speaker 2: still servicing a very large number of clinics around the
8:51
Speaker 2: world providing agent clock aging biomarker services. And we also
8:57
Speaker 2: iPod and Silico last year in the set so on
9:00
Speaker 2: the thirtieth of December were listed and kind of the
9:04
Speaker 2: last day of the year and managed to pull the
9:08
Speaker 2: largest fire tech IPO in Hong Kong in twenty twenty five,
9:12
Speaker 2: so then we've celebrated the new year, and since then
9:16
Speaker 2: the performance of the company has been very respectable, so
9:20
Speaker 2: we won on doubled. However, I still think that we're
9:23
Speaker 2: significantly undervalued.
9:25
Speaker 1: I haven't met a thinking. I don't know any other
9:28
Speaker 1: CEO of a biotech company who thinks they're overvalued.
9:32
Speaker 2: I know a few.
9:35
Speaker 1: Oh, I'd love to meet them. No, but of course
9:37
Speaker 1: you believe in your own effort. And I have to
9:40
Speaker 1: say this sounds I'm going to the next time I
9:42
Speaker 1: see you, I'm going to bring you a book, the
9:44
Speaker 1: biography of Michael Bloomberg, who's obviously the founder of the company.
9:47
Speaker 1: You'll find some parallels in there in that he financed
9:50
Speaker 1: the early days of Bloomberg completely with his own money.
9:53
Speaker 1: And I think there was a six month angle there too,
9:55
Speaker 1: at some point where if there was no deals coming,
9:59
Speaker 1: he would have run out to cash. So I think
10:01
Speaker 1: it's that it was six months a year. I don't remember,
10:04
Speaker 1: but it's in the book. I'll bring it to you
10:06
Speaker 1: as a present, please do so. There are some good
10:08
Speaker 1: parallels there. The interesting thing, of course, here, Alex, is
10:12
Speaker 1: that you have done quite a lot of work in
10:15
Speaker 1: I think, can I call it the call face of
10:17
Speaker 1: actually using AI to discover drugs and doing it fully right,
10:21
Speaker 1: So I'm sure there's an element of human You know,
10:23
Speaker 1: you've got a medicinal chemist as your co CEO, and
10:26
Speaker 1: we'll come back to that. What I'm what I'm interested
10:28
Speaker 1: in is if you could just tell me a little
10:30
Speaker 1: bit about position in Silicon for us in terms of
10:32
Speaker 1: how many people you are where you're based. I do
10:36
Speaker 1: know that you have some Middle East presence in terms
10:38
Speaker 1: of engineers and data center. If I'm not wrong, just
10:42
Speaker 1: tell us. Just just paint that picture of in Silicon
10:44
Speaker 1: for us right now.
10:45
Speaker 2: Sure, So in Silico is a trull, truly global entity.
10:48
Speaker 2: We now have multiple regional sites performing very specific tasks.
10:54
Speaker 2: I'm Canadian by citizenship and also Latvian, so I could
10:58
Speaker 2: have chosen many places to be here, but I really
11:01
Speaker 2: love Canada. That's why we also have in Silicon Montreal,
11:05
Speaker 2: where we developed models for AI drug discovery and work
11:10
Speaker 2: on a frontier foundational modeling. We have a site in
11:13
Speaker 2: Cambridge in Boston, which is predominantly focused on clinical development
11:17
Speaker 2: and business development IR and we have a site in
11:22
Speaker 2: Abu Dhabi where we scale AI and we have more
11:25
Speaker 2: than sixty people there, also focused on material science and sustainability.
11:30
Speaker 2: And we have a site in Hong Kong where we're
11:32
Speaker 2: listed where the Hong Kong Science Park, and that's where
11:35
Speaker 2: we discover targets. As you know, Hong Kong is kind
11:37
Speaker 2: of like the Switzerland of China, and some of the
11:40
Speaker 2: people working at Silico Hong Kong they work there for
11:44
Speaker 2: more than five years. So we are very focused on
11:47
Speaker 2: novel target discovery there. And we have sites in Shanghai Pudn.
11:51
Speaker 2: We have two sites, one on John Jong another one
11:54
Speaker 2: is right next door in Dutch and Park where we
11:56
Speaker 2: have a fully automated facility that allows us to run
12:00
Speaker 2: experiments at scale, mostly in target discovery, but also an
12:05
Speaker 2: indication expansion safety. We'll also have a dry lab and
12:09
Speaker 2: we are currently developing a protein therapeutic so biologics lab
12:14
Speaker 2: in yieshing So City, right next to Sujio. It's about
12:17
Speaker 2: maybe two and a half three hours drive from Shanghai,
12:21
Speaker 2: so that's our global footprint. Basically, we develop AI in
12:26
Speaker 2: Montreal and Abu Dabi where we scale and then we
12:28
Speaker 2: utilize AI to validate and test in China and other countries.
12:35
Speaker 1: Well, so we're very fortuitous today that it is very
12:39
Speaker 1: fortunate that is the day after you see announced you
12:42
Speaker 1: deal with Eli Lilly, And we'll come back to that
12:44
Speaker 1: in terms of interactions with large forma, et cetera. But
12:47
Speaker 1: let's go from top down here. If you wanted to
12:51
Speaker 1: characterize where AI is today in drug discovery, drug development,
12:58
Speaker 1: what can it do? What coun't it do? It is
13:02
Speaker 1: obviously since GPT came out into the world, world's consciousness
13:06
Speaker 1: in twenty twenty two, early twenty twenty two, everyone a
13:10
Speaker 1: lot of people think that that was when AI was invented,
13:12
Speaker 1: which of course is not the case. But perhaps just
13:16
Speaker 1: tell us where you think AI sits today in the
13:19
Speaker 1: world of discovering to the spectrum, what can it do?
13:24
Speaker 1: What can't it do?
13:26
Speaker 2: So that is a very good question. AI can do
13:28
Speaker 2: a lot. I think it already can perform most of
13:32
Speaker 2: the tasks that are required to discover a drug. And
13:36
Speaker 2: when I'm talking about Ai'm talking about multiple different models,
13:40
Speaker 2: multiple different algorithms that can perform specific functions and also
13:45
Speaker 2: orchestration algorithms foundation models that kind of are called as
13:50
Speaker 2: agents in order to orchestrate and reason within the entire
13:56
Speaker 2: therapeutic drug discovery and development program, where AI help you
14:00
Speaker 2: the most in drug discovery is help you go from
14:04
Speaker 2: zero to the point, which we call preclinical candidate, develop
14:09
Speaker 2: or development candidate. There are multiple terminologies, but let's refer
14:14
Speaker 2: to this one step before clinical trials as PCC. And
14:19
Speaker 2: different companies have different standards for pccs. Our standard is
14:25
Speaker 2: for a given protein target, in order to nominate the PCC,
14:29
Speaker 2: you need to have multiple in vitro well. First of all,
14:32
Speaker 2: binding studies showing that your moleculative designed binds to the
14:37
Speaker 2: protein of interest. Then you demonstrate cellular potency, so you
14:42
Speaker 2: test it in different cell types cell models. Then very
14:46
Speaker 2: often we go organoid and tested on organoids. Then you
14:49
Speaker 2: go different types of animals for efficacy studies. So whatever
14:54
Speaker 2: animal models you've got you need to test it. Usually
14:56
Speaker 2: you use it use two or more animal models of disease.
15:00
Speaker 2: Then we perform what we call non GLP toxicity studies.
15:06
Speaker 2: Those are twenty eight day non GLP toxicity experiments in
15:11
Speaker 2: at least two species. For in competitive target space where
15:15
Speaker 2: we don't focus on target novelty but focus on molecular novelty.
15:19
Speaker 2: We have to do sometimes four species, and we also
15:23
Speaker 2: synthesize every single competitor that we can get hands on
15:27
Speaker 2: the structure. Very often we have to guess those molecules
15:30
Speaker 2: from patterns, and once the package is complete, we nominate
15:34
Speaker 2: the PCC. It is kind of like five PhD defenses
15:38
Speaker 2: can compressed into one because of course you've got opposing
15:42
Speaker 2: parties that want to ensure that the molecule complies with
15:46
Speaker 2: the highest standards because they need to license those molecules out.
15:50
Speaker 2: Like for example, I'm also on the BD side, so
15:54
Speaker 2: I am always criticizing the molecule from the business development perspective,
15:59
Speaker 2: and after you reach PCC, that's it. So that's where
16:03
Speaker 2: acceleration that you can get from AI stops. After that
16:07
Speaker 2: you move with speed of traffic because when you start
16:10
Speaker 2: doing your I and D enabling studies, those are usually
16:14
Speaker 2: twenty eight day GLP texicity studies in two species, so
16:19
Speaker 2: you go mouse and dog, or mouse and monkey, or
16:23
Speaker 2: rat and dog, rat and monkey, depending on what is
16:27
Speaker 2: common for the target you are prosecuting and the molecular
16:32
Speaker 2: class you are using for your program. So those are
16:36
Speaker 2: highly regulated experiments. That's why it's called GLP good laboratory practices.
16:41
Speaker 2: In order to conduct those GLP experiments. The lab must
16:46
Speaker 2: be certified in the US and Europe and other geographies
16:50
Speaker 2: to comply with many different protocols and procedures and regulations.
16:55
Speaker 2: So it actually costs you a lot of money and
16:58
Speaker 2: usually it takes about nine to two twelve months to
17:01
Speaker 2: just repeat what you did in non GLP tox fashion
17:04
Speaker 2: in four months or less because those are twenty day studies.
17:09
Speaker 2: But you are doing that in a regulated manner. And
17:12
Speaker 2: after that phase one, phase two, phase three human clinical trials,
17:16
Speaker 2: that's where you really move with a speed of traffic.
17:19
Speaker 2: There is no way to accelerate, and you can only
17:21
Speaker 2: use AI for regulator submissions, for reviews, et cetera, et cetera.
17:26
Speaker 2: So currently I can perform many, many, many, many different tasks. However,
17:31
Speaker 2: you still need to perform experiments in order to see
17:34
Speaker 2: if AI worked or not and create those high quality
17:37
Speaker 2: data rooms that pharmaceutical companies would require in order to
17:42
Speaker 2: inlicense your AI generated drug, because for them, it doesn't
17:46
Speaker 2: matter how you discover the drug. You could have used magic,
17:50
Speaker 2: you could have used a genie, you could create you know, well,
17:56
Speaker 2: sheare serendipity or just brute force. Right now, most of
18:00
Speaker 2: my competitors are in China, and many of them are
18:03
Speaker 2: using brute force, and they don't care. As long as
18:06
Speaker 2: the molecule has a great data room and you're out
18:09
Speaker 2: performing the standard of care and other competitors, the molecule
18:13
Speaker 2: may be transactable. So we need to ensure that the
18:16
Speaker 2: data rooms contain all information and have all the checkmarks
18:20
Speaker 2: that big Pharma needs in order to buy.
18:22
Speaker 1: So let me let me see if I've understood you correctly.
18:25
Speaker 1: Let me summarize here to go from concept to lead nomination,
18:31
Speaker 1: which I guess is post GLP ts right, how much.
18:36
Speaker 2: Phenomenal candid candidate domination? Okay, So we first do lead optimization,
18:43
Speaker 2: after that candidate domination, and after that we do GLP tals.
18:47
Speaker 1: Okay. So to the point of starting GILP talks, what
18:50
Speaker 1: is how would you classify the difference that an AI
18:54
Speaker 1: generated process or AI supported process gives you versus the
18:59
Speaker 1: brute force process if you have access to broot force.
19:02
Speaker 2: So in small molecule space where we compete, and that's
19:07
Speaker 2: by the way, much more hardcore than on biologics because
19:11
Speaker 2: you need to worry about toxicity much more. But then
19:14
Speaker 2: small molecules give you superpowers because it's a much more
19:18
Speaker 2: scalable therapeutic modality, and also you can ship it around
19:23
Speaker 2: the world, store it for a long time, make it cheaply,
19:25
Speaker 2: and give it to emerging geographies. So in small molecules,
19:29
Speaker 2: usually a big pharmaceutical company using traditional approaches would synthesize
19:35
Speaker 2: thousand or more molecules. In the past, when we just started,
19:38
Speaker 2: like two thousand and eighteen ish, they were synthesizing two
19:42
Speaker 2: and a half thousand per program. Just to give you
19:45
Speaker 2: a or put you in perspective your first date with
19:48
Speaker 2: a molecule. So synthesize, synthesis and tests would cost you
19:52
Speaker 2: around five thousand bucks, well two and a half to
19:54
Speaker 2: five thousand, depending on how good your fts A the
19:57
Speaker 2: cro are, And that's just the b assays. If you
20:01
Speaker 2: want to synthesize more and test and more exotic assays,
20:05
Speaker 2: you are going to pay much more than that. So
20:09
Speaker 2: multiply that by two and a half thousand, and then
20:13
Speaker 2: you're left with a few hundred. You optimize, you test again.
20:18
Speaker 2: The costs compound With AI. At Insilico, we usually synthesize
20:24
Speaker 2: sixty to two hundred molecules, so that's basically an order
20:28
Speaker 2: of magnitude delta between our traditional approaches. Plus, since why
20:33
Speaker 2: using generative metal methods instead of searching for a needle
20:37
Speaker 2: in a haystack, we can very rapidly generate a bunch
20:40
Speaker 2: of needles. So those sixty to two hundred molecules already
20:44
Speaker 2: have significant diversity of scaffold in order for us to
20:48
Speaker 2: explore the chemical space and try many combinations permutations, and
20:52
Speaker 2: then we just nuddel down to one scaffold, optimize just
20:57
Speaker 2: a little, and you get a drug, right.
21:00
Speaker 1: Either to quantify the so there's two things you've understood
21:02
Speaker 1: here is the number of potential leads that you generate
21:05
Speaker 1: can be magnitudes or does the magnitudes higher? But also
21:09
Speaker 1: the speed with which you generate them, if you wanted
21:11
Speaker 1: to put to quantify the speed, how much faster are
21:15
Speaker 1: you getting to your potential lead versus standard methods.
21:19
Speaker 2: So traditional approach with then big pharmaceutical company outside China,
21:25
Speaker 2: usually the time to develop mutal candidate or with a
21:29
Speaker 2: moderately novel target would be around four and a half years,
21:32
Speaker 2: sometimes four years, because they will still rely on China.
21:35
Speaker 2: Within the MNC multinational company in China, they have a
21:39
Speaker 2: local R and D center and go after law to
21:43
Speaker 2: moderate novelty targets. They can get away with about two
21:48
Speaker 2: and a half years, still longer because many processes are bureaucratic.
21:52
Speaker 2: At in Silico, our typical timeline are between nine to
21:56
Speaker 2: eighteen months, so that's a pretty substantial reduction. But can
22:02
Speaker 2: we also compete most of the time nowadays on timelines
22:06
Speaker 2: on molecules, on the program quality, program level quality with
22:11
Speaker 2: the local Chinese companies, and those are not just like
22:16
Speaker 2: AI drug discovery companies or smaller biotechts. Those are usually
22:20
Speaker 2: large companies like you know Hungary, like Fosso and Chilus, CSPC,
22:26
Speaker 2: his Son and many many others. So they learned how
22:30
Speaker 2: to scale and some of them have over one hundred
22:32
Speaker 2: programs in different stages of discovery and development. Some of
22:36
Speaker 2: them managed to go to PCC in like twelve months
22:40
Speaker 2: with biologics, sometimes even shorter timelines, so very very short timelines,
22:45
Speaker 2: and then they actually know how to progress in the
22:46
Speaker 2: clinic and there it's extremely difficult to compete, but we
22:51
Speaker 2: still compete on those timelines, And where we compete in
22:56
Speaker 2: most of the time is actually not only the time,
22:58
Speaker 2: speed and cost, even though they are, we still can
23:02
Speaker 2: outperform traditional Chinese pharma companies, but we mostly compete on novelty.
23:08
Speaker 2: So traditional Chinese companies they usually do not go after
23:12
Speaker 2: either novel targets or dramatically novel scaffolds. And actually they
23:18
Speaker 2: don't need to because big pharmaceutical companies that are nowadays
23:22
Speaker 2: are quite eager to replenish their pipelines. They do not
23:26
Speaker 2: care as much about novelty. They hear about confidence so
23:31
Speaker 2: that the program will reach the clinic that it will be,
23:35
Speaker 2: it will be successful, and once it succeeds, it must
23:38
Speaker 2: outperform the standard of care. So very often if this
23:41
Speaker 2: is your objective function and also potentially blockbuster, if that's
23:46
Speaker 2: your objective function, you are much better off going after
23:50
Speaker 2: something that is a low novelty target, or you're targeting
23:54
Speaker 2: two or more very well known known targets with a
23:59
Speaker 2: biologic like an ADC or like you know, do all targeting,
24:03
Speaker 2: try targeting antibodies and get significant patent life on the
24:07
Speaker 2: competitive advantage in the US and European markets and make
24:12
Speaker 2: a lot of money. So they are not exactly interested novelty,
24:15
Speaker 2: and that's where Chinese companies thrive. So moderate novelty or
24:19
Speaker 2: low novelty target moderate novelty to low novelty molecule at
24:24
Speaker 2: low price with high confidence. That's what China wants. Sorry,
24:28
Speaker 2: I was what pharma wants from China. So at in Silico,
24:32
Speaker 2: we can compete on novelty, so we can go after
24:36
Speaker 2: higher novelty targets, like some of my targets have never
24:38
Speaker 2: been in the clinic before. That's very rare to go
24:41
Speaker 2: after those kind of targets at scale because you can
24:45
Speaker 2: only make a few bets on. Pharma doesn't want to
24:48
Speaker 2: buy them at before phace to complete, and sometimes even
24:52
Speaker 2: that is not enough. They would like to get an
24:55
Speaker 2: improved molecule, or they want to see you progress even further.
25:00
Speaker 2: That's risky, and AI gives us confidence to go after
25:03
Speaker 2: those targets because we know that if we establish the
25:06
Speaker 2: president and we also have very kind of we're not
25:11
Speaker 2: a monopoly, but we are a super expert and a
25:13
Speaker 2: specific target class category with multiple backups with a highly
25:18
Speaker 2: protected chemistry chemistry pattern space. They would like to partner
25:22
Speaker 2: with us and give us a lot of money. But again, unfortunately,
25:27
Speaker 2: even on the kind of the current deals that we
25:30
Speaker 2: have and we have made many of this year. Most
25:33
Speaker 2: of the companies want to have low novelty on a target,
25:36
Speaker 2: high novelty on a molecule, and that's sorry. I can
25:40
Speaker 2: help a lot.
25:40
Speaker 1: That's interesting because if you recall, I did this thing
25:43
Speaker 1: trying to get a feel for innovation in China, looking
25:46
Speaker 1: at the number of possible first in class molecules that
25:49
Speaker 1: could come to market by twenty thirty with some obvious assumptions.
25:54
Speaker 1: What I hadn't looked into is to try and classify
25:58
Speaker 1: these molecules or any fun pipeline a novelty because that's
26:03
Speaker 1: that's difficult. It depends a lot on your definitions. But
26:08
Speaker 1: let's let's talk about this subject. So the the target's
26:11
Speaker 1: biology identifies genes are thrown up in the biology or
26:16
Speaker 1: in the system or somehow. The chemistry is also an
26:19
Speaker 1: extra angle here. Now you fish in the known chemical space.
26:24
Speaker 1: I understand, I suppose how or correct me there.
26:28
Speaker 2: How we actually fish. So we have two ponds, both
26:35
Speaker 2: with high novelty high novelty entities. So in one pond
26:41
Speaker 2: we fish for targets. So we have biological AI biology
26:46
Speaker 2: for you two that allows us to prioritize targets. Very
26:51
Speaker 2: often when we're doing it for sale. To ensure that
26:56
Speaker 2: we keep the lights on. We go after a lot
27:00
Speaker 2: novelty targets or modern novelty targets because we know that
27:03
Speaker 2: that's what pharma will want to buy first.
27:06
Speaker 1: You mean, can you tell me what that would be?
27:08
Speaker 1: Is that's something a molecule that wasn't known or characterized
27:12
Speaker 1: before in a pathway.
27:13
Speaker 2: Or well, think about like keiras. Okay, if we all
27:17
Speaker 2: know the ke ras protein because it has been drugged,
27:22
Speaker 2: it's uh, there are proof drugs or keiras, and it's
27:26
Speaker 2: kind of a Kardashian of ancology targets. Everybody knows it.
27:31
Speaker 2: If you don't know that, well, you're from different culture.
27:35
Speaker 2: And or like GLP one Okay, like everybody knows GLP
27:39
Speaker 2: one because South Park made an episode on that.
27:43
Speaker 1: So that's low novelty. But you can come up with
27:46
Speaker 1: a better drug with better chemistry.
27:49
Speaker 2: Yeah, so you basically compete on the chemistry side. So
27:52
Speaker 2: usually for most of those high value targets, somebody already
27:57
Speaker 2: developed a molecule or a billo, so they've already tested
28:01
Speaker 2: the target on the clinic and it's very well known,
28:03
Speaker 2: very well known to work. However, there's always room for
28:07
Speaker 2: improvement on the molecular side. So the way you drug it,
28:09
Speaker 2: you can either make drugs safer so reduce some liabilities,
28:14
Speaker 2: and that might open some new indications, like for example,
28:17
Speaker 2: a drug that you would normally take for cancer can
28:20
Speaker 2: all be used for a CNS disorder beautiful or like
28:24
Speaker 2: autoimmune so something that you actually need to take for
28:27
Speaker 2: your entire life to modulate your immune system specific dose.
28:31
Speaker 2: If it's super safe, but it was used for cancer before,
28:34
Speaker 2: you may be able to repurposent additional indications and you
28:38
Speaker 2: just redesign the molecule or design something absolutely from scratch
28:42
Speaker 2: for that target that we know where biology works. And
28:45
Speaker 2: there you still risk a lot because like small modifications
28:50
Speaker 2: on the molecular scaffold of the molecular targets that hits
28:55
Speaker 2: the target can also result on pretty dramatic effects and
28:59
Speaker 2: also site effects that you would not be able to predict.
29:02
Speaker 2: So your AI has to be very good at molecular
29:05
Speaker 2: design and then program level reasoning. So however, the main
29:10
Speaker 2: pond that where we fish at in Silico will actually
29:13
Speaker 2: go and discover new targets, new mechanisms, So something that
29:17
Speaker 2: Big Farma never saw working in the clinic. So imagine
29:21
Speaker 2: that the next GLP one and they are all searching
29:23
Speaker 2: for the next GLP one. So how do we find
29:26
Speaker 2: next GLP one? Well, we have to we have AI
29:30
Speaker 2: for that, and we have many forms of AI for that.
29:32
Speaker 2: And when many of those AIS agree on the target
29:36
Speaker 2: for a specific disease. And also I often go for
29:39
Speaker 2: targets that are dual purpose implicated on aging and disease
29:43
Speaker 2: at the same time. Kind of my favorite target category.
29:46
Speaker 2: Those are really targets with potential for pipeline a product
29:50
Speaker 2: where you can approve for one indication and go into
29:52
Speaker 2: many indications. And we have many of those. But if
29:56
Speaker 2: you bat on one of those targets and selected for
29:59
Speaker 2: your own pipeline, you need to be prepared to go
30:02
Speaker 2: phase two or at least phase one too complete with
30:05
Speaker 2: limited number of patients. Because for novel targets, pharmaceutical companies
30:10
Speaker 2: either public partner very early and do it kind of
30:13
Speaker 2: enough form of collaboration, so they give you some money
30:16
Speaker 2: and then phase some targets, they look at them and
30:18
Speaker 2: almost never progress. I have not seen a single successful
30:22
Speaker 2: target discovery partnership in my life, and that includes us,
30:25
Speaker 2: by the way, because very often pharma finds great targets,
30:29
Speaker 2: we deliver those targets, but then they have strategic reprioritization
30:34
Speaker 2: and they still decide to bat on more more proven
30:38
Speaker 2: stuff internally, or they're chased chasing the next big trend
30:41
Speaker 2: and switching from oncology to obesity. So target discovery is
30:48
Speaker 2: a dead end from the commercial perspective. You can make
30:51
Speaker 2: one or two partnerships on target discovery, but that's it.
30:55
Speaker 1: So for novel targets, you need to take it through
30:57
Speaker 1: to proof of concept decision in the phase two exactly, okay.
31:01
Speaker 1: And then in the chemistry space in this small molecule world,
31:05
Speaker 1: what took us through that? You know, the chemical space
31:09
Speaker 1: is very large, the combinate combinatorial world is very large.
31:13
Speaker 1: How does how much new chemistry do you feely identify that?
31:19
Speaker 1: Just quantify that for us space.
31:21
Speaker 2: Sure. So there are two ways to prosecute the target
31:25
Speaker 2: you selected with generative AI and small molecule chemistry. So
31:30
Speaker 2: one way is there might be already a starting point
31:34
Speaker 2: that is very well known, either an own molecule that
31:38
Speaker 2: is that target. Sometimes it might be actually a drug,
31:41
Speaker 2: and you can take that scaffold kind of that molecule
31:45
Speaker 2: as a starting point and try to redesign it. So
31:47
Speaker 2: imagine that you want like a perfect well I don't
31:52
Speaker 2: want to use human Imagine that you want to have
31:55
Speaker 2: a perfect path and your friend has a dog that
31:59
Speaker 2: you like, but that dog is barking too loud at night.
32:03
Speaker 2: So now you've figured you just need to figure out
32:05
Speaker 2: how to get a similar dog that doesn't bark and
32:08
Speaker 2: as patentable when it comes to molecule molecules, but it's
32:12
Speaker 2: going to be very similar. That's something that you like
32:14
Speaker 2: very much. But another approach is you know the target
32:18
Speaker 2: that you want to go after, and there is a
32:20
Speaker 2: crystal structure, or you make a crystal structure, or in
32:23
Speaker 2: a case of alpha Wold, you can actually predict the
32:26
Speaker 2: crystal structure and unfl that crystal structure. You give your
32:29
Speaker 2: generative AI software the pocket where small molecule might bind.
32:34
Speaker 2: So imagine there is no chemistry available. You have to
32:37
Speaker 2: start from scratch, and you can give it a pocket,
32:40
Speaker 2: or you can ask AI to automatically identify pockets, and
32:44
Speaker 2: then you generate molecules that potentially bind this pocket or
32:48
Speaker 2: several pockets, and at the same time have multiple beneficial
32:52
Speaker 2: drug like properties. So not just the molecule it binds,
32:56
Speaker 2: but also the molecular binds, and it's likely to make
32:59
Speaker 2: a very good So those two approaches both are tractable
33:04
Speaker 2: and can be performed by I, and we have done both.
33:08
Speaker 2: So one is similar to something existing and patentable. And
33:13
Speaker 2: by the way, pharma loves that too. They actually don't
33:16
Speaker 2: really need to have absolute molecular novelty. Sometimes they think
33:20
Speaker 2: that it's a liability because with something that is similar
33:24
Speaker 2: but pattern three with improved properties, you are more likely
33:28
Speaker 2: to not see some unexpected effects because somebody else already
33:33
Speaker 2: has proceeded into the clinic or through the clinic with
33:36
Speaker 2: something similar. And to answer your question in terms of
33:39
Speaker 2: capabilities in terms of exploring the chemical space. So in
33:43
Speaker 2: both cases, in the first one it cannot go too
33:46
Speaker 2: far from the original scaffold because otherwise it will defit
33:50
Speaker 2: the purpose. But with a second approach, where doing crystal
33:53
Speaker 2: based design pocket or pocket based design, you can actually
33:56
Speaker 2: be very creative and ask your AI system to generate
34:00
Speaker 2: molecules that are you know, less than zo point one
34:04
Speaker 2: motive similarity to the original, or they might not be
34:08
Speaker 2: original at all.
34:09
Speaker 1: So Alex, in the broader AI software world, there's been
34:14
Speaker 1: this recent question arising as to how which of the
34:19
Speaker 1: software models out there that the legacy type software models,
34:24
Speaker 1: software as a service, or even some of the more
34:26
Speaker 1: advanced point of sales software companies, a whole variety, which
34:31
Speaker 1: one can survive the onslaught of open source models and
34:37
Speaker 1: even just a whole lot of these very large LLM
34:41
Speaker 1: companies that we all know open AI anthropic that they
34:46
Speaker 1: could come and just eventually displaced you, partly by providing
34:51
Speaker 1: replicating what you do, or partly by people using their
34:54
Speaker 1: their capabilities developing their own What is the what is
34:57
Speaker 1: the risk to I don't want to say in city specifically,
35:01
Speaker 1: but companies that use AI for developing and discovering drugs,
35:05
Speaker 1: when do you expect it to become a commodity in
35:07
Speaker 1: five years, ten years or whatever time for you?
35:10
Speaker 2: A very good question. So, first of all, actually think
35:14
Speaker 2: that if you were to do just pure AI drug
35:16
Speaker 2: discovery software and AI for science and then with intend
35:21
Speaker 2: to sell it to bigger players, I think you would
35:24
Speaker 2: not survive anyway. So that is not a very survivable
35:28
Speaker 2: place because you need to invest a lot in order
35:31
Speaker 2: to stay irrelevant. But the market is very small, so
35:34
Speaker 2: people analysts always overestimate the size of the aipower AI
35:40
Speaker 2: drug discovery software market, so where you can actually sell
35:44
Speaker 2: software a platform and I think it's billions, but I
35:46
Speaker 2: actually think it's not even a billion. I think it's
35:49
Speaker 2: maybe a few hundred million. And you can tell the
35:52
Speaker 2: technology investors all kinds of stories. Usually those are the
35:56
Speaker 2: ones that like platform companies, but they lack experience and
36:00
Speaker 2: drug discovery, so they don't understand how the industry works.
36:03
Speaker 2: They will ask those companies or what kind of proprietary
36:06
Speaker 2: data you have, what kind of talent you have, what
36:09
Speaker 2: compute you have, But in reality, none of those investments
36:13
Speaker 2: work for many reasons. Well one of them just the
36:16
Speaker 2: market is too small. And if you are doing your
36:19
Speaker 2: own drug discovery, you're actually very well protected, especially if
36:23
Speaker 2: you are going after a more or less novel target
36:26
Speaker 2: or target space, because people will value you just on
36:29
Speaker 2: the pipeline. They will not give you a huge credit
36:32
Speaker 2: for being in an AI company. Sometimes they do not with us,
36:35
Speaker 2: or not with any companies that I am friends with.
36:38
Speaker 2: And once you have a clinical well, once you have
36:41
Speaker 2: a preclinical click and a clinical ass portfolio that provides
36:44
Speaker 2: you with tangible well first of all evidence that you
36:47
Speaker 2: Arai works, but also with a portfolio of drugs that
36:51
Speaker 2: you can go and license out, so it gives you
36:54
Speaker 2: some confidence and also can generate some internal revenue. So
36:58
Speaker 2: that's why I'm saying, usually two investors who want to
37:01
Speaker 2: go late stage with a company that doesn't have many
37:04
Speaker 2: pccs or has like log Qualley, I'm recommending just you know,
37:08
Speaker 2: don't invest in the kind of venture rounds because it's
37:12
Speaker 2: not sustainable and at some point in time they will
37:14
Speaker 2: have to do their own drugs and if the caverns
37:17
Speaker 2: started yet, it might be already too late.
37:20
Speaker 3: Right.
37:20
Speaker 1: So the proof of value in in Silico and other
37:24
Speaker 1: companies potentially, like you said, I don't know if there
37:26
Speaker 1: are others, but I'm sure you will tell me, is
37:28
Speaker 1: that the of the value proposition is number one. We
37:33
Speaker 1: go after in silic, who goes after novel targets themselves
37:38
Speaker 1: that folks haven't targeted before, at least publicly announced, and
37:43
Speaker 1: we have the ability to test the chemical space very
37:47
Speaker 1: broadly and cleverly and get it. So in the end,
37:51
Speaker 1: that's that's what biotechs do. A lot of them start
37:53
Speaker 1: with some science and technology or chemistry or or drug
37:59
Speaker 1: design that they had worked on or discovered within a
38:01
Speaker 1: lab and then they brought it out and here you're
38:05
Speaker 1: recreating that lab too. So I've got that right, Have
38:09
Speaker 1: I summarized that? Right?
38:10
Speaker 2: Oh? Correct? Okay, Yeah, there are kind of several ways
38:13
Speaker 2: to do that. And on silica we try to go
38:15
Speaker 2: after higher novelty stuff on both target and small molecule side. However,
38:21
Speaker 2: often we have to be opportunistic and go after lower
38:24
Speaker 2: novelty targets, like moderate novelity targets, just because pharma wants it.
38:28
Speaker 1: Yeah, we all have to live with the reality of
38:30
Speaker 1: what's out there, but at the same time try and
38:32
Speaker 1: innovate to a point where, you know, with your bigger
38:34
Speaker 1: picture view.
38:34
Speaker 2: Right.
38:35
Speaker 1: So let me bring this to the point that I
38:37
Speaker 1: think maybe you and I discussed a little bit when
38:39
Speaker 1: we saw each other, is you don't know what you
38:41
Speaker 1: don't know? Is it possible that some of the novel
38:44
Speaker 1: targets that you look at, we're actually tested in some
38:47
Speaker 1: lab in a GSK or Advisor or Lily or one
38:51
Speaker 1: of these big, big places which has a ton of
38:53
Speaker 1: work going on pre clinical that never sees the light
38:56
Speaker 1: of day. And of course if they've failed or because
38:59
Speaker 1: they shot down those projects that never come out. Does
39:02
Speaker 1: that matter even if that's the case. And then added
39:05
Speaker 1: to that, do you feel that you have a disadvantage
39:09
Speaker 1: not having access to those massive data sets that are
39:13
Speaker 1: in house within large pharma. Just talk about that subject
39:17
Speaker 1: a little bit, please, sure.
39:18
Speaker 2: So, very good question, by the way, And if you
39:22
Speaker 2: are to post this podcast on some pharma related medium, well,
39:27
Speaker 2: it's actually a good idea to highlight it. Highlight this
39:30
Speaker 2: section because what we realized over time is that many
39:33
Speaker 2: targets that we are focusing on that are absolutely novel.
39:37
Speaker 2: Pharma has tried them before, not all, but some and
39:42
Speaker 2: they've tried to go after that target, spend several years
39:46
Speaker 2: on it, and then the prioritize and very often you
39:49
Speaker 2: come to a scientist to sell the molecule and you
39:54
Speaker 2: show the program and they're saying, oh, when I was
39:57
Speaker 2: at this and this company, I was actually working on
40:00
Speaker 2: the target. We worked on this for two years later
40:02
Speaker 2: abandoned the project and yeah, actually think it's too difficult,
40:06
Speaker 2: et cetera, et cetera. That's why they make such valuable
40:09
Speaker 2: business development people, right so, because they know what doesn't
40:14
Speaker 2: didn't work and what people worked on in general in
40:16
Speaker 2: addition to other things, and very often we see that
40:21
Speaker 2: they actually got to a point where it was very promising,
40:24
Speaker 2: but due to strategic reasons they decided to deprioritize, or
40:28
Speaker 2: because some famous scientists loved the company or decided to
40:33
Speaker 2: focus on something else. So that's very common, right however,
40:37
Speaker 2: and we can only know whether they worked on this
40:40
Speaker 2: target or not by looking at the patents. So if
40:44
Speaker 2: they have worked on a target seriously, usually they would
40:48
Speaker 2: have generated a bunch of molecules that they also decided
40:54
Speaker 2: to patent for PCC, usually at PCC or just before that,
40:59
Speaker 2: and you would you would normally worked on that, but
41:02
Speaker 2: very often you won't. So that the punchline here is
41:05
Speaker 2: that very often, most often they don't know that either
41:10
Speaker 2: because because the scientist has left or the target got
41:13
Speaker 2: the prioritized, et cetera. And you know, just a few
41:17
Speaker 2: years passes and they don't even remember if they worked
41:19
Speaker 2: on a target or not, and if they did. Usually
41:23
Speaker 2: the data rooms that are created for exploratory research, these
41:28
Speaker 2: are not production grade data rooms. Uh, these are very
41:32
Speaker 2: limited data rooms because they are doing exploratory research for
41:36
Speaker 2: the purpose of trying to identify some certain depitous biological
41:39
Speaker 2: process process, and very often it doesn't really or result
41:45
Speaker 2: in really good data. And they just described the pro
41:48
Speaker 2: the program and the entire data set.
41:51
Speaker 1: Right, So Alex a huge So if we don't have
41:54
Speaker 1: a huge amount of time left and I've got a
41:56
Speaker 1: huge number of questions, still, let's let's talk about think
42:00
Speaker 1: we've really explored well the idea of and the use
42:04
Speaker 1: of AI up to the pre clinical candidate stage. So
42:09
Speaker 1: I was listening to Patrick Schwab who's at gsk AI,
42:13
Speaker 1: and he was talking about the value or the impact
42:16
Speaker 1: of AI on the clinical development part. So there's two
42:19
Speaker 1: questions I'm going to load in here. One is the
42:22
Speaker 1: FDA wants to get rid of animal testing in pre
42:24
Speaker 1: clinical so and it wants to replace it by AI.
42:28
Speaker 1: Your view on that, because you obviously said that's something
42:31
Speaker 1: that I don't think we can replace it. We have
42:33
Speaker 1: to go through it, et cetera. And then the Patrick
42:36
Speaker 1: Schwab white space. Going from phase one to phase two
42:38
Speaker 1: to phase three, we've talked about the time saving and
42:41
Speaker 1: the value of AI in discovery and to the PCC point,
42:45
Speaker 1: talk to us about these two things, first, replacing the
42:48
Speaker 1: preclinical models with AI, and two and how much would
42:51
Speaker 1: that happen and how what's your view and how much
42:53
Speaker 1: time will it save? And secondly, the clinical development path
42:56
Speaker 1: which you are into, now can you not cut? Of course,
42:59
Speaker 1: if you need to do a twelve month human study,
43:01
Speaker 1: I suspect you have to do it to a month
43:03
Speaker 1: human study, but more rapidly finding patients, better drug targeting
43:08
Speaker 1: in terms of identifying the patient. I should go in
43:10
Speaker 1: there cutting the white space between finishing phase one going
43:13
Speaker 1: to phase two. Talk to us about that, please and
43:16
Speaker 1: see if at the end we can quantify how much
43:18
Speaker 1: time you could or couldn't save from the clinical development
43:21
Speaker 1: side and pre clinical side.
43:23
Speaker 2: Right well, on the pre clinical side, as I've already explained,
43:27
Speaker 2: you can cut considerably in terms of time. So we
43:31
Speaker 2: cut to nine to eighteen months from zero to PCC's
43:35
Speaker 2: compared to traditional methods, and not by the way, includes
43:40
Speaker 2: already twenty eight day non GLP texasity studies and animals.
43:43
Speaker 2: So on. GLP talks were just repeating the same thing,
43:46
Speaker 2: and my first answer, can we replace animal testing? No, Okay,
43:52
Speaker 2: we'll be able to do it in five years. Cut
43:55
Speaker 2: animal testing now? Okay, great, ten years maybe, but I
44:01
Speaker 2: don't think so. Okay, it's just in order for you
44:04
Speaker 2: to sell those drugs later to pharma companies. You actually
44:07
Speaker 2: need to get this data anyway, and you actually are
44:12
Speaker 2: very optimistic about your molecule and about your EI until
44:15
Speaker 2: you see some dogs die because of unexpected lever or
44:19
Speaker 2: kidney talks, usually liver because they get like very weak livers.
44:23
Speaker 2: By the way, if dogs respond poorly to a drug,
44:26
Speaker 2: it doesn't necessarily mean that humans will. It's a bad
44:29
Speaker 2: sign and it's not going to let you kind of
44:31
Speaker 2: sell a drug easily. But usually humans tolerate drugs much
44:37
Speaker 2: better than dogs. But it's important to have this data. Second,
44:41
Speaker 2: so white space, I think that right now pharma companies,
44:45
Speaker 2: all of them want you FACE to complete assays, sorry assets,
44:49
Speaker 2: and for many drugs where we wherever we are going
44:54
Speaker 2: after like novel targets, they would all want to have
44:57
Speaker 2: FACE two and they don't want to partner early. For
44:59
Speaker 2: NOVELI and all novelty targets, they would be able to
45:03
Speaker 2: partner very early.
45:05
Speaker 1: But in terms of cutting the time that one year
45:07
Speaker 1: that it takes from going from the end of phase
45:09
Speaker 1: two to the start of phase three, do you feel
45:11
Speaker 1: there is some space for AI to help you speed
45:14
Speaker 1: that up? Document production.
45:15
Speaker 2: I think it's going to be very minor. So the fact,
45:20
Speaker 2: once you reached PCC, it's kind of like you already
45:24
Speaker 2: have a gun. So you build a gun and you
45:26
Speaker 2: build a bullet, and at PCC stage you fire a bullet.
45:31
Speaker 2: So after that the bullet is traveling, you can adjust
45:34
Speaker 2: the aim a little bit, you can blow wind on
45:38
Speaker 2: it if you can, you can do many other things,
45:41
Speaker 2: but you can choose the you know what to shoot at,
45:46
Speaker 2: like the patient selection. But in reality, there is not
45:51
Speaker 2: that much you can do to the bullet. And in
45:53
Speaker 2: terms of the speed you can help with regulatory submission,
45:57
Speaker 2: with the regular submission, regular submission, review response. That's why
46:01
Speaker 2: I actually sometimes helps you correct human errors and I've
46:04
Speaker 2: seen that before or well it can hallucinate too, but
46:09
Speaker 2: it actually does a pretty good job at generating those
46:13
Speaker 2: bureaucratic documents. However, the cost of submission, meaning in terms
46:17
Speaker 2: of the decision cost, right, So imagine if you submit
46:20
Speaker 2: something and you're running the risk of THEDA not giving
46:24
Speaker 2: you the go ahead. So the cost of that for
46:28
Speaker 2: a biotic is so huge that you would not only
46:30
Speaker 2: not use AI at scale, of course, you would do
46:34
Speaker 2: it at the very beginning for templating, but later you
46:36
Speaker 2: would also invite a bunch of high profile experts to
46:41
Speaker 2: help you review the regulatory documents.
46:45
Speaker 1: Right, okay, so let me summarize again. The time and
46:47
Speaker 1: cost saving comes in the very early parts of drug discovery.
46:51
Speaker 1: The rest of it's some minor in terms of the
46:53
Speaker 1: development timelines. Probability of success. Where do we stand on that?
46:57
Speaker 1: Do you feel that you should would be able to
47:00
Speaker 1: increase the probablative success to get to successful phase three
47:05
Speaker 1: having generated you molecus with AI and the a help
47:09
Speaker 1: of AI. Think?
47:10
Speaker 2: So?
47:11
Speaker 1: Okay? I think so if you were to give me
47:13
Speaker 1: a number, give me a number, I mean just for conversation,
47:18
Speaker 1: right that that if if if a phase three has
47:22
Speaker 1: a say, a sixty percent probabative success when it starts,
47:25
Speaker 1: where would where would you hope to be with the
47:29
Speaker 1: application of good AI once you've got.
47:31
Speaker 2: The Well, here you go, so we have at a
47:34
Speaker 2: silico so far we've never failed. Think about that? So
47:39
Speaker 2: in the like past, non.
47:42
Speaker 1: GI mean, once you've got into the clinic so.
47:44
Speaker 2: Far even an I D enabling Right, So that's why
47:47
Speaker 2: I'm actually I think that from the theoretical perspective, you
47:51
Speaker 2: couldn't theory drop I D enabling and go directly into humans. Right,
47:55
Speaker 2: So you actually don't need to drug those wonderful naive
47:58
Speaker 2: animals because out of the twelve twelve times we push
48:01
Speaker 2: the drug into the clinic, which means that we have
48:04
Speaker 2: passed at least twelve I in the enabling stages, non
48:08
Speaker 2: failed and I in the enabling non failed on phase
48:11
Speaker 2: one so far, non failed in phase two. And when
48:16
Speaker 2: I'm saying that, I'm actually extremely worried, right because at
48:20
Speaker 2: some point in time you need to fail. Just statistics
48:23
Speaker 2: tell you that you have like zero points zero zero
48:27
Speaker 2: zero zero zero one percent chance of being where we
48:29
Speaker 2: are right now.
48:30
Speaker 1: Yeah, but that's just those statistics are derived from the
48:35
Speaker 1: history of drug development to date. I mean, I'm not
48:38
Speaker 1: defending it. I'm not saying that you won't have a failure.
48:40
Speaker 1: I have no idea. But at the end of the day,
48:42
Speaker 1: that's based on what we've been doing so far going forward.
48:45
Speaker 2: Yeah, So I'm not trying to promote the company here,
48:48
Speaker 2: I'm just saying that, I'm just saying that in our case,
48:52
Speaker 2: I just have only one source of evidence, my company,
48:58
Speaker 2: that I can use to reason and make educated gases
49:03
Speaker 2: because nobody studied this properly. Actually, I think BCG and
49:08
Speaker 2: also some academics tried.
49:11
Speaker 1: Yes, but do you know how much human intervention there
49:14
Speaker 1: was in that?
49:15
Speaker 2: Yeah, I discuss because so if you look at some
49:18
Speaker 2: of those reports that like BCG put out and Nature
49:22
Speaker 2: Use Drug Discovery I think, or some other big journal
49:26
Speaker 2: I think it was an RDD, they have like one
49:29
Speaker 2: hundred plus molecules in different stages. And if you look
49:33
Speaker 2: at the companies that produce those molecules, some of them
49:37
Speaker 2: started in like two thousand and two or something like this,
49:41
Speaker 2: and most of them I don't know as AI players.
49:45
Speaker 2: They just kind of rebranded. So some of those molecules
49:49
Speaker 2: are very old, and still they managed to demonstrate that
49:52
Speaker 2: even with that you know, kind of ancient wave of
49:57
Speaker 2: quote unquote AI drug Discovery company, there is some improvement
50:01
Speaker 2: in probability of success. But so far, I think they
50:05
Speaker 2: showed like one or two of phase three completes of
50:09
Speaker 2: some kind of unknown dinosaurs that completed the studies and
50:13
Speaker 2: some of them repurposed. Right, Okay, last question, actually yeah, yeah.
50:18
Speaker 1: Last question. Even though it's paining me that you have
50:21
Speaker 1: to stop here, I wish I was like, there are
50:23
Speaker 1: podcasts that go on for two or three hours. But
50:25
Speaker 1: I think Alex is going to just turn the camera
50:28
Speaker 1: and the phone off and disappear. Let's let's talk about
50:32
Speaker 1: the latest Deerio did, which.
50:33
Speaker 2: Was a great time my case. I'm sorry actually, like
50:36
Speaker 2: you know, the entire weekend on the week prior, I
50:40
Speaker 2: was sleeping two three hours. So even right now I'm
50:45
Speaker 2: a little kind of dizzy.
50:47
Speaker 1: Well you're doing really well for someone who's dizzy, not
50:50
Speaker 1: at all. It's perfect. We you know, this is not
50:52
Speaker 1: our normal conversation about AI in drug discovery. So I'm
50:56
Speaker 1: glad you took it slow with us. But tell us
50:58
Speaker 1: about this latest deer of which of course is with
51:01
Speaker 1: the largest farmer company in the world, one of the
51:03
Speaker 1: largest R and D budgets. So how long did it take?
51:07
Speaker 1: And of course it seems to me that you were
51:11
Speaker 1: pushed or rushed a little bit to get the press
51:13
Speaker 1: face out because of market regulations and rules. Just talk
51:17
Speaker 1: to us whatever you can tell us in the openly
51:21
Speaker 1: about this deal.
51:23
Speaker 2: So sure, this deal is a combination of an asset
51:27
Speaker 2: licensing and collaboration. We've collaborated with Lily in the past
51:31
Speaker 2: and they are actually one of the most active and
51:34
Speaker 2: expert users of our software, so they deployed the entire
51:37
Speaker 2: pharmaire suite. That's why I usually referred to Lily as
51:41
Speaker 2: being better than us at AI because they actually have
51:44
Speaker 2: all the lowries and they also know how to use it.
51:48
Speaker 2: So a phenomenal company with absolute experts on the biologics
51:55
Speaker 2: and the small molecule side. And what I really love
51:58
Speaker 2: about Lily is the that they actually don't have not
52:01
Speaker 2: invented me, not invented here attitude. They actually kind of
52:06
Speaker 2: try to fight it. Like once I talk to Andrew Adams,
52:10
Speaker 2: the head of neuroscience and also head of genetic medicine,
52:14
Speaker 2: truly phenomenal scientist and a very young pharmaceutical executive who
52:19
Speaker 2: truly has vision and experience, and I said, well, you know,
52:22
Speaker 2: one of your molecules have not been designed by you,
52:26
Speaker 2: and that's why very successful. So do you really here?
52:30
Speaker 2: He's like, I don't care. As long as the molecule
52:33
Speaker 2: works on the clinic. The source of the molecule doesn't
52:36
Speaker 2: really matter. What matters is that we get it approved,
52:40
Speaker 2: get it to the patient, and issure to deliver benefit
52:44
Speaker 2: and outperform everything else. I'm like, wow, that's an interesting attitude,
52:49
Speaker 2: so I usually deeply care that you know, it's my drug,
52:53
Speaker 2: and that actually gives made me thinking that maybe I
52:56
Speaker 2: should go and then license some drugs because nowadays I
53:02
Speaker 2: know the entire market and I know the competition, synthesize everything.
53:06
Speaker 2: So Lily, when we started partnering, they already knew us,
53:12
Speaker 2: but they realized that we have some assets they want.
53:15
Speaker 2: And once that happened, we actually some of those assets
53:20
Speaker 2: are in a huge demand in the pharmaceutical industry, so
53:24
Speaker 2: highly competitive. So several companies looked at, like identified those assets,
53:30
Speaker 2: started looking at them, and some of them started placing bids,
53:35
Speaker 2: and all of them were substantial, like very large. So
53:40
Speaker 2: once once Lily placed the bid, I just started racing.
53:44
Speaker 2: I decided that, you know, that is the company I
53:49
Speaker 2: love for this type of deal is just like an
53:54
Speaker 2: idea of pharma. And I mean, there's well, yeah, but
54:00
Speaker 2: they trade plus minus Fizer every month, right, So if
54:08
Speaker 2: I don't want to offend anyone, no numbers numbers such
54:12
Speaker 2: a giant and also for a good reason. It's not
54:15
Speaker 2: just the good lucky with one drug, it's it's consistency.
54:19
Speaker 2: And we thought that you know, if we partner with
54:23
Speaker 2: Lily on those ultra high value, high demand programs. Well,
54:28
Speaker 2: we managed to pull a feet of molecular engineering that
54:33
Speaker 2: is absolutely incredible. It's like selling you know, ground complication
54:39
Speaker 2: watch to Rusher on Constantine, So it needs to be
54:44
Speaker 2: absolutely amazing. So we have something like that, I mean
54:48
Speaker 2: we had, and the deal took. They managed to pull
54:52
Speaker 2: it very quickly. It's basically four months.
54:55
Speaker 1: Probably I think that's a record. Now. I mean, I'm
54:58
Speaker 1: sure there'd be deal right faster than that, but four
55:00
Speaker 1: months is not credible.
55:02
Speaker 2: I must say, like, I don't want to disclose anything
55:05
Speaker 2: like that so that we don't set the precedent for others.
55:10
Speaker 2: Because I also must say that I've seen that in China,
55:14
Speaker 2: but I have never seen something like that on the
55:17
Speaker 2: US and on the global level, where people worked, people
55:20
Speaker 2: worked extraordinary hours and also you know, really focused on
55:26
Speaker 2: the deal. And the business development team and the transaction
55:31
Speaker 2: team at Lily is second to none. I've never seen
55:36
Speaker 2: anything like that. Is just absolute focus, dedication, professionalism, and
55:42
Speaker 2: maybe we got lucky, most likely we did, but just
55:46
Speaker 2: the level of gear, level of professionalism, level of diligence,
55:52
Speaker 2: so they managed to check mark all the boxes and
55:54
Speaker 2: the way go Way Beyond because they are experts in
55:58
Speaker 2: some of the spaces where we partnered all of them,
56:01
Speaker 2: all of those spaces, it's basically like a gun like
56:05
Speaker 2: rusher on Constantine evaluating your ground complication to beyond system.
56:11
Speaker 2: If you know watches and I don't know watches, I
56:14
Speaker 2: just looked at I'll bring out there and they managed
56:22
Speaker 2: to check mark all the boxes very quickly and also
56:26
Speaker 2: add additional levels of collaboration where we pass additional expertise
56:30
Speaker 2: to them and partner become a partner of the ecosystem
56:34
Speaker 2: called Lily Gateway Labs, and they've just recently constructed, so
56:38
Speaker 2: it's kind of a more advanced version of Johnson and
56:42
Speaker 2: Johnson's g Labs where you create a community the ecosystem.
56:48
Speaker 2: So we're happy to be part of that. And if
56:51
Speaker 2: anyone is expanding into China and wants to get a
56:55
Speaker 2: great platform to start with, Lilly Gateway Labs is an
56:59
Speaker 2: amazing place is to do that and also for the
57:03
Speaker 2: Chinese companies to become more global. So those Gateway Labs
57:08
Speaker 2: just an incredible advance and you know, I don't think
57:13
Speaker 2: it's humanly possible to do something like that. Again, it's
57:16
Speaker 2: just I think you cannot do it faster than four months.
57:19
Speaker 1: No, probably not, given the.
57:22
Speaker 2: Usually those deals take like nine months.
57:24
Speaker 1: This also includes the Chinese New Year and the Christmas
57:27
Speaker 1: if I calculate correctly.
57:28
Speaker 2: So, yeah, that's why I don't want to say a
57:31
Speaker 2: thing about kind of you know, somebody can point the
57:34
Speaker 2: finger and say or unfair labor practices or whatever, but
57:38
Speaker 2: as you know, you know, cancer and aging do not
57:42
Speaker 2: take holidays holidays. Yeah.
57:44
Speaker 1: Yeah. But also Alex, you have a business that's spread
57:47
Speaker 1: across you know, into the you know, a w aw
57:50
Speaker 1: you don't have Christmas there, you don't have the Chinese
57:52
Speaker 1: New Year there. But anyway, so I'm just generally.
57:54
Speaker 2: So and I try to escape holidays whenever I look
57:59
Speaker 2: at myself because as my job is to deliver drugs
58:02
Speaker 2: to patients that die otherwise.
58:04
Speaker 1: Well, we really appreciate the time that you've spent trying
58:07
Speaker 1: to help us and our audience to think through your
58:10
Speaker 1: perspective of in AI and Silicon's perspective in AI. So
58:13
Speaker 1: thank you very much, Alex. So that's Alex Javoronkov, the
58:17
Speaker 1: founder and CEO co CEO of Ellen Silico Medicines. Thank
58:21
Speaker 1: you so much for joining us for our latest episode.
58:24
Speaker 1: I have a whole lot more questions. I'm going to
58:26
Speaker 1: keep them from when I see you. Maybe we have
58:28
Speaker 1: another dinner. Thank you, Sam, Thank you. Please don't forget
58:31
Speaker 1: to give us a review and make sure to click
58:33
Speaker 1: the follow button on your favorite podcast app or site
58:37
Speaker 1: so you never missed a discussion with the leaders in
58:40
Speaker 1: healthcare innovation such as the one we just had.
58:42
Speaker 2: Absolutely will do.
58:44
Speaker 1: Thank you, Sam Fazeli, You've been the listening to Vanguards
58:47
Speaker 1: of healthcare podcast by Bloomberg Intelligence. Until next time, take.
58:51
Speaker 3: Care USUS and as pass US uses some of the
59:39
Speaker 3: jases