The Harry Glorikian Show

How Drug Development Guru Mark Eller Went from AI Skeptic to AI Supporter

Episode Summary

How does an expert in pharmacokinetics, whose only exposure to computers was taking one semester of programming in college to meet a language requirement, become an advocate for the new AI-driven style of drug discovery? This week Harry finds out from Mark Eller, who helped to invent Allegra at Hoechst Marion Roussel (now Sanofi), spent 12 years at Jazz Pharmaceuticals; and is now senior vice president of research and development at twoXAR, an AI-driven drug discovery startup.

Episode Notes

How does an expert in pharmacokinetics, whose only exposure to computers was taking one semester of programming in college to meet a language requirement, become an advocate for the new AI-driven style of drug discovery? This week Harry finds out from Mark Eller, who helped to invent Allegra at Hoechst Marion Roussel (now Sanofi), spent 12 years at Jazz Pharmaceuticals, and is now senior vice president of research and development at twoXAR, an AI-driven drug discovery startup.

In our previous episode from August 31, 2020, Harry spoke with twoXAR founder and CEO Andrew A. Radin, who confessed to being a computer nerd and lamented that it's been hard finding colleagues who are willing and able to help him bridge the gap between software and biology. He told the story of Mark Eller, who started out as a consultant at twoXAR but ended up telling Radin "I want you to offer me a job." 

Eller told Radin that the twoXAR project had finally convinced him that AI is good for more than just winning games of chess or Go, and that it can also be used to help drug developers predict which new molecules will be effective against specific diseases, even if their mechanisms of action are unfamiliar.  "He's gone from highly skeptical to highly supportive," says Radin. "I think that transformation is happening throughout the industry." This week, Harry gets the whole story of Eller's transformation, from Eller himself.

Please rate and review The Harry Glorikian Show on Apple PodcastsHere's how to do that from an iPhone, iPad, or iPod touch:

1. Open the Podcasts app on your iPhone, iPad, or Mac.

2. Navigate to The Harry Glorikian Show podcast. You can find it by searching for it or selecting it from your library. Just note that you'll have to go to the series page which shows all the episodes, not just the page for a single episode.

3. Scroll down to find the subhead titled "Ratings & Reviews."

4. Under one of the highlighted reviews, select "Write a Review."

5. Next, select a star rating at the top — you have the option of choosing between one and five stars.

6. Using the text box at the top, write a title for your review. Then, in the lower text box, write your review. Your review can be up to 300 words long.

7. Once you've finished, select "Send" or "Save" in the top-right corner.

8. If you've never left a podcast review before, enter a nickname. Your nickname will be displayed next to any reviews you leave from here on out.

9. After selecting a nickname, tap OK. Your review may not be immediately visible.

That's it! Thanks so much.

Harry: All right, Mark. Welcome to the show.

Mark: Thank you very much for asking me, huh? 

Harry: Yeah, no, it's great to have you here. Um, and I, and I honestly, I'm really excited about this conversation after Andrew mentioned, like, you know, he sort of gave me a sneak peek of, of, uh, The story. And I was like, we gotta do the show.

Mark: Like we gotta have this conversation. 

Harry: And so Mark, just, just for, you know, everybody that's listening, sort of, can you give us a little bit about, of your background because you're, you're not the, you know, computer science, uh, born and bred person, right? That a lot of the people that I will have on the show you're. You're hardcore drug discovery

Mark: I mean, I had one programming course in college that I actually took to fulfill a foreign language requirement because, you know, programming languages were the neat new thing and they wanted to encourage that. So yeah, I have no sort of. Technical knowledge in programming and my background, my training is in clinical pharmacology and, uh, basically spent the last 30 years in the drug industry in big pharma and CRO, uh, on.

Nonclinical preclinical development and clinical pharmacology PKPD. And, um, you know, before I started twos, I was consulting for a while. And before that I was at Jazz were 15 year or not 15 feels like, like 12 years, uh, starting out as VP of research and then had various titles along the way. And along the way I became inventor on something like that 30 patents, uh, including the ones, uh, the original ones for the anti-histamine, uh, Lycra, and then at, uh, uh, Jazz on some of their, uh, Xyrem patents and, um, just, uh, you know, the spent my time in developing drugs. That was my passion. And what I'd like to do, especially the early stages of drug development.

And then started, uh, consulting after I left Jazz and twoXar was one of my first clients. So that's how, sort of, how I got to, uh, got to them., 

Harry: So so essentially, you know what I would consider, you know, a traditional drug development, uh, background. Well, I mean, Obviously distinguished, like you've done some incredible stuff cause my son takes Allegro. So I'm sure he's very happy that you developed it. Um, but you know, I guess your first exposure to this was dealing with twoXar as a client. If I understand correctly.

Mark: Yeah, that's fine. So, you know, they were, um, mostly. Uh, computer and bioinformatics people. They had a few people with drug development background, but they needed, they realized they needed more in that area.

And through a friend of a friend, they reached out to me and I was, uh, consulting. And so I said, sure, I'll be happy to help them. And you know what I knew about. AI at that time was like from documentaries. I had seen doc, you know, computers beating chess, and then there was a one you a more recent one about alpha go and about the masters and the team from the, you know, Google bought them and they've got the. They're playing the master and the, uh, computer program wins and they're oohing and nine about this and they can't figure out why the computer, you know, made this winning move or why it came up from there. Yeah. And you know, that was my understanding of AI and I couldn't understand how AI could possibly help with drug development, because I could see, okay, for a game, I can see how like a computer could play against itself over and over again, and figure out what, you know, what are the winning moves and learn to think ahead and things like that.

But for drug development, you know, if you use that game analogy and if you use like, um, Success as an FDA approval, then there's only a few thousand games that have ever been played. And so how is the computer supposed to iterate on that? And if you define an approval and approved drug as a win as the target that you're going after, then how are you ever going to come up with a better drug or a drug for a disease where there, you know, there aren't drugs approved, which is the really.

The goal. So I, I didn't understand how AI could help with that. But as I became consulting with them, you knowthey obviously, they didn't bring me in for the AI part. They brought me in to say, okay, we find, you know, we found this interesting molecule and we have these, how should we test it? Or we've tested it.

And now we have these results, help us help us with that or help us with the next step. And, you know, I would do that and I would say, Oh, No, these, this looks pretty good. This is neat. I went to off the bat night happened, but obviously it did. And then sort of retrospect, they can figure it out. Okay. So now we should do this, you know, and I would see that.

You know, that that pattern would repeat itself where they would, you know, they would run their computer platform and come up with like nine or 10 molecules and put them in a nonclinical model of efficacy and two or three of them would pop out positive and the, the hook for me was they all had new mechanisms of action.

So these were mechanisms of action that hadn't been tested in the patient population for that disease. And so, you know, that was an immediate challenge to my understanding of AI. So how did the computer come up with. That. And I didn't know the answer to that question, but you know, the more and more I saw results like this, the more I'm more, I thought, however, it's doing it.

It's, he's producing results. Centering interesting, you know, at a, sort of a higher success rate than a lot of, uh, Uh, traditional methods of, of drug discovery. And that was something that was very appealing to me and I wanted to be a part of, so I started, uh, talking to Andrew then and, uh, sort of ended up as senior vice president of R &D here.

Harry: So what, you know, walk me through. Sort of ha you know that from skepticism to, you know, I want to be part of this it's, you know what you did, you said, look at you. We showing you. 

Mark: It was the highways. It was the results from the past, but, but also my understanding of, of AI sort of changed. So, you know, I joined in, uh, November of last year, 2019, and in September of 2019, I was reading this article by Bob temple, who is, uh, this senior person at the FDA. And he's been around for a long time. He was involved with, uh, the approval of Allegra and he was talking about how there was this new age in drug development can call it the age of individualization or like personalized medicine.

Right? So recognizing that individual differences in patients, uh, contribute to individual variations in response. And they're there, you know, the, he called the previous ages, the ages of safety and efficacy, and he was citing regulation and he said, but for this age there was, it wasn't really kicked off by regulation.

It was kicked off by the discovery of drug interactions with Seldin and I was involved in doing those. Drug interaction studies. And it was actually those studies that led to the invention of Allegra, which didn't have the, you know, the problems that led selling to be with withdrawn from the market. But it, it struck me that what was happening then was like the very beginning of the introduction of signal detection into the pharmaceutical industry. You know, cell being on average was very safe, uh, so safe that the company wanted to go OTC. And that's what I was hired in to do, to help them repeat some studies and get it ready for OTC. But then we got started getting these reports of drug interactions and, uh, arrhythmias.

And it was the first sort of application of signal detection in the pharmaceutical industry signal detection signals from individual patients. So looking at individual patients for four signals, and at first, you know, the, the signal detection system was so crude that the only signal that broke through were arrhythmias. And then we learned how to. Parse a peat, a piece of the ECG, the QT interval as a signal detection system. And, you know, as a result of those, uh, drug interactions, FDA also introduced MedWatch. Which is a signal detection system to get adverse event data from marketed drugs. And then, you know, that's one of the things, FDA monitors, how the drugs are actually working in the marketplace and are there any unexpected reactions and things like that.

And yeah, it's sort of hit me then that what twoXar was doing and what, how they were using AI was actually as a signal detection system, sifting through genomics data and phenotypic data and all these different datasets and examining it with an AI system to look for efficacy signals. And with that sort of reframing of.

You know, drug discovery as a problem in signal detection, it makes it a problem that's amenable to a computational solution and allows you to apply, you know, methods of signal detection that have been used in other places, other industries. Uh, to pharmacology and. You know, the results that I, then it sort of clicked and it fit with the results that I was saying, Oh yeah, you know, the computer wasn't playing a game to come up with this compound.

It was sifting through this data looking for signals. And then, um, you know, we tested it and lo and behold, the signal was, was validated at least in the, the first, uh, animal model or in vitro test. And so with that sort of reframing of my understanding of AI along with sort of. Re sort of defining the drug discovery problem.

It started to all fit together. And that's when I thought, you know, wow. You know, if signal detection 20 years ago, 25 years ago, according to Bob Temple was the thing that ushered in this new. Era in, you know, development or pharmaceutical development. Well, if we apply signal detection to the beginning part to efficacy, you know, I think it just, it, it has tremendous potential and I'm just, uh, it's a price that hasn't been done before because it's really kind of, uh, the results are better than let's say just a much better than I would have anticipated going into this.

Harry: Now, so it wasn't just necessarily the results, but I think if I heard you correctly, it's also this factor of, time being shorter

Mark:. Yes. It was. It's much more efficient. I mean, you know, Tuesday was a very small company and yet, you know, they've got 18 programs lunch and I think on 10 of them, uh, we have in vitro or, uh, animal pharmacology data with positive results and, you know, It would have taken years and years, especially for a little company, like, like this, to generate that data with traditional methods.

So it's like, Much more efficient. It gets you out of the triangle that manager's always talking about, you know, cost and time and quality. And you can get any two of those, but you have to sacrifice on, on that third one. Well, this just sort of, you know, breaks that all open and you can get really.

Good results, uh, are very fast and you know, much more efficiently than the, than the traditional approach. So it was, it was just a big jump, um, all the way around from the traditional way of doing things. 

Harry: Now, the other thing though, that you said is, and so I'm trying to use these two axes, right? One is time. One isyou are seeing patterns or pathways that. You were like, ah, not seen that before and opening up sort of a unique area to look at that you might be able to develop a new molecule again, which of course is a. 

Mark: Yeah, exactly. So that was, that was the really the exciting part. So, you know, there's a lot of, a lot of drug development and a lot of drugs are like there, you know, me too, drugs, they're second generation or third generation of something that does this and maybe there's incremental improvements, you know?

And the third statin that it's approved as better than the first one or whatever, but they're working by the same mechanism. What was interesting to me is that, you know, twoXar’s approach and their platform would identify molecules that were potentially effective in a given disease. And the approved ones would show up to or stuff that might be in development by the companies would show up to, uh, which, you know, it gives you some comfort that you're on the right track.

But what also show up is. Stuff that not only that hadn't been tested, but, but that have mechanisms where there's no drug that's approved or even been tested in patients with that disease that have that mechanism. So it's a whole new way of sort of treating the disease. So it's like the computers come up with a new or the actions of the computer, the whole system, yes.

Equal included have come up with a, like a new, a new hypothesis for what might work to treat a given disease that, that hasn't been tried before. And then we test those in animal models and it would. You know, test 10 of them in three of them pop up, uh, effective. That was very, very exciting. 

Harry: Yeah. I've, you know, I've talked to a lot of people where they have it, you know, they have a well understood process of making a molecule and then they, you know, to their system and say, well, you know, is there a faster way, more efficient way? And the system can sometimes tell you how to get to the same end result in a different way than anybody was classically trained.

Uh, that might bring down cost and decrease time and so forth. And so this doesn't sound, you know, I do can see similarities between all these approaches. Um, now the other thing though, that when I talk to Andrew, he's like, you know, his hypothesis was, you could shave off three to four years of time. In the whole process. Like, do you agree with that sort of, yeah.

Mark: I mean, from, uh, what we've done so far from, you know, the, the, uh, things that we've taken furthest along in and have results back from, uh, nonclinical pharmacology models. Uh, yeah. You know, it was like something like four months. From, you know, saying we want to investigate this disease to having results in a laboratory animals.

And, you know, doing them at CRO is the same. CRO is at big pharma uses in sort of standard classical models. And most of that time, frankly, was, uh, for the actual conduct of the study, not the, all the, um, stuff that. Brought us to that prediction and the stuff that, you know, You know, traditional process that I would bring you to, that prediction might take, might years.

And here, you know, we had collapsed all those initial steps down into one that just generates efficacy predictions that can be, you know, immediately tested. So, so yeah, there's that sort of time and efficiency aspect of it. And then, you know, when there are new molecules or new mechanisms of action that made it really, really exciting.

Harry: So now, I mean, but at some point the system generates what it is. It suggests. And then the human being though has got to look at it and be like,

Mark: Yes of course. Yeah. Okay. So, you know, I can give you like a flow step and the first step is like all the computers. And then the last step is basically, you know, humans deciding what animal models should we use to test this.

Hypothesis. And then the intermediate steps, you know, span the range of more computer and less human to more and more human. So, yeah, uh, after the initial, you know, generation of possibilities, which might be, you know, A thousand come out or something like that. Then there is this winnowing process, uh, and ranking process.

And the first few steps of that are also computer assisted or AI assisted to rank them or, uh, eliminate molecules that might be too toxic or for a given indication, things like that. And there there's human intervention along the way until finally the decision as to which of the molecules to put into a non-clinical model is, uh, dependent on human insight and, uh, not, not AI.

Harry: So, but at some point that information now you are working with the people actually building the models, right? So there's yes. I assume that, that, you know, I think of this as a figure eight, right. It just, at some point there's feedback, there's correction or there's modification of the, and then it just keeps going back and forth then just makes it better.

Mark: Yes, exactly. 

Harry: I mean, it was interesting cause I was talking to someone earlier. Today from a big pharma and saying, you know, one of the companies I was talking to has said, they're constantly improving their, their algorithm. And he says, now we don't do that at big pharma because, you know, our model is pretty well set. And, and, you know, we're, unless we think that there's going to be some huge breakthrough. 

Mark: I know, I think we're on, I don't know. This is a question for Andrew, but version. I don't know, but I I'm sure. Pretty sure it's in, you know, hundreds of iterations of the, um, stop where, yeah. So there's this constant learning process.

Harry: So do you, do you see this actually also affecting cost in development? 

Mark: Well, I think, you know, cost and timing sort of go hand in hand. It really, the way you achieve the tiny is because you're eliminating all these, you know, wet lab rate, limiting experiments, and those wet lab rate-limiting experiments also cost money. So yeah, there's, there's a big, um, cost savings as well as, as well as a time saver. 

Harry: But yeah, I think about it from a, uh, uh, uh, rate limiting, but also. I think the number of parameters, these algorithms can look at it as much more than, you know, definitely, definitely more than me. Right. 

Mark: I mean, that is how it comes up with new stuff that you and I haven't or other people haven't thought of.

It's, you know, I think, um, for one of the, uh, programs, um, lupus, I think, uh, they were like, Uh, Tom was talking to Erin who already, who heads up there, runs the platform. And it was something like 2.5 billion pieces of data that, uh, were going into this, that it was sifting through. So yeah, it's, it's more than more than humans can handle.

Harry: So, but, and then I always think about like, the papers that have to be coming out of this. I mean, at some point, you know, you got to start to. Let the world know that there's this other potential pathway that you could use this, you know, or, or, you know, just to publish this stuff and say, here's a different way to come at these problems and make it more of a widespread now I know, you know, you know, as a startup, you want to be first and own it all. But I think about that just from a science perspective.

Mark: Oh yeah. Yeah. I agree. I mean, there is, you know, there's that tension between proprietary and getting the information out. And so in terms of results, That's certainly something, uh, that we are publishing, for example, the lupus one on, I know that the abstracts have been presented or, you know, all the meetings are virtual now.

Uh, yeah. Uh, yeah, that where we've presented the, the data, uh, from those studies and also a couple of others, like, uh, cellular caution, Noma. And so the data is starting to come out and people can, you know, Judge for themselves and take a look at it's very positive. 

Harry: So now, uh, for, if you were talking to other people like with that, that have your background, so what, you know, how, how do you, and I'm sure it's come up in, in you've I'm sure you have this conversation with, with, uh, colleagues.

Yeah. It's what do you say to them and how do you frame it in a way that they sort of can get there? Their head around it, um, quickly, if right. And, and, and what are the skeptical comments that you get? Cause I'm sure that you and I get probably the exact same comments. 

Mark: Yeah, I think it depends upon, you know, it depends, it depends upon what they're looking at. So if you're, if they're looking at the results of a particular program, if they're looking at our SLD data, I just say, you know, just, just look at the data and decide for yourself and then ask yourself, does it really matter? You know, if I came up with the idea at work and Peter came up with it, so there was this, this game that I used to.

Um, play at Jazz. We were doing like, uh, opportunity assessment and we called it doing a pre-mortem. So we pretend, okay. We, we bought this asset and we did everything and executed according to plan and it failed. Why did it, why did it fail? And if you write down all the reasons it could potentially fail.

And you know, the last reason is because the computer came up with the idea. Yeah. That wouldn't stop us from moving forward with that program because in the end, you know, the FDA doesn't care. If a computer came up with the idea or, you know, whoever it, they care about a development plan about, um, development, rationale and your data.

So I, it's very easy to convince people on individual products. I think on. You know, the AI side, it's, it is more people are more skeptical and they're more like in my position and I have to say, look it, you know, put away your preconceptions about what AI is and you know, if you're like me and you thought it was about computers, playing games against humans and they, you know, it's a little bit, you know, maybe that's part of it, but it's a little bit more than that.

And if you view it as signal detection, Um, that might help you understand what the, what the potential is. And, you know, for me, that helped and also seeing the results, you know, repeat over and over, you know, convinced me that that was, uh, that was in fact the case. But yeah, we do run into people who are skeptical and say, Oh yeah, you know, this data looks really nice. And that piece of data looks really nice and this program looks good, but you know, they're not. I guess convinced that there's a block there and they can't see that, uh, you know, if it works three times or five times or 10 times that it's likely to, you know, keep working and, you know, we think that there are, you know, a thousand diseases that it might be applicable for.

There's some probably that it's not, you know, like infectious diseases, it's the platform. It's not set up to do something like that because you've got a third thing, the organism itself. But, uh, for other things, it's yeah, it's very broadly applicable. 

Harry: Yeah. I mean, uh, you know, we've always talked about it. Identifying new targets. Completely different pathways that, you know, have a major effect on the disease that nobody ever considered. Right. Um, and then, uh, the third is, is repurposing something that's already out there that might actually have a meaningful effect on this particular disease, but nobody is nobody's really using it in that way.

Mark: Right. Yeah. Yeah. You know, sometimes we come up with stuff like that too, you know, we'll be in our list of things that are predicted to be potentially effective. There might be, you know, a drug that's, um, on the market for something else. So that does happen occasionally. 

Harry: So, so, if you were coaching somebody and say, and that was working their way through the system. Would you tell them to study something? Would you tell them to. I don't know, read a particular book. Would you, how would you sort of coach someone along that was the you're younger? You? 

Mark: Um, I would say, always be guided by science and, you know, look at the results and look at the hypothesis and connect the dots. And if, you know, if you connect the dots and it says signal the texting AI works, you know, trust that and believe it and put away your preconceptions. You don't have to understand the details of the programming to see the results. And if you can just understand the inputs, and if you can understand that the idea that the computer can take two and a half billion pieces of data and process it, and you can't do that, um, And comes up with predictions.

Well, test the predictions. It's sort of empirical it's, shouldn't involve, you know, I'm a scientist, it doesn't involve belief or trust or whatever. It's like, here's the, here's the output. Did the output work? Yes, here's the input and this experiment here's out, but did it work? Yes. How many times do you need to, to see that to, to be convinced and, you know, you can just sort of be agnostic in your, uh, beliefs about AI or computers or traditional approaches and, and just, you know, just because you're used to doing something one way, And it's worked, although maybe, you know, slow and inefficient, you know, don't be closed to the idea that there might be other ways to do things that, uh, get you out of that triangle of cost and time and quality.

Harry: Yeah. Another thing that I think about is the systems are it's, it's not like they're standing still either. They're in a constant state of improvement and evolution, which is just making them better over time. And we are collecting more data sets that the system can ingest and build into the model. So I think it's just, it's moving forward at a, at a pretty, at a very fast pace. I mean, you know, I try, I try to explain the, the speed at which things are happening to people. And it's very difficult for the human mind to understand. Doubling doubling, like has a ha I don't know why we have a hard time getting around.

Mark: I mean, it's like, you know, science, sometimes things in science moves sort of in a linear, predictable pattern. And sometimes it's just, it jumps like, you know, the, cell vein drug interactions leading to the MedWatch system and signal detection for safety. What that, that didn't come up out through a linear process.

That was a jump and a response to a problem. And the solution sort of developed Denovo. And this is sort of, I think, a similar situation, it represents sort of a break from the traditional way of doing things. And you, you know, if you're open to an objective to looking at results, you should be sort of.

Okay with that. Um, you know, but there, you know, we get comments from, you know, you know, where's your, you know, five KOL’s in this disease area who came up with this thing, you know, the whole purpose of this, you know, we're a small company and we've got, we've got stuff in. That we're developing for SLE. We've got it for oncology indications. We've got some, you know, you can partnered with, uh, people in, in various therapeutic areas. We don't have, you know, we have, we hire a KOL as consultants and we need them, but we don't have this, you know, staff of therapeutic disease experts who have. You know, working for 20 years to come up with this molecule, but you know, there there's, you know, to give people credit.

There's a lot of companies in the Bay area that that's how they are start to some professor from Stanford and working on this problem for 20 years. And then they spin it off into a little startup company. And people can, you know, wrap their head around that, whereas, okay. Uh, Andrew and Aaron were working on this computer system and now we're going to apply it to pharmacology and we can apply it to oncology.

We can imply it, uh, inflammatory diseases and we can apply it in other areas and, you know, we've got results and that just, just it's different. 

Harry: Yeah, yeah, yeah. Yeah. I mean, I think in the other part is, is. If some new pathway comes up, that the system came up with. I mean, I'm, I'm hard pressed to find a KOL that really like, you know, would have gone. Oh, Oh yeah. That was, we absolutely knew about that. Right. It's what I find is they're like, Oh, I mean, I didn't even think about that one. And yeah, now that you're showing me the data that's short of, I can see how that might make sense. It just wasn't. Right in front of them. 

Mark: Yeah. It's what I would call retrospective predictability.

Now I can't think of it, but now, now that now that the computer did and all that, you've got positive results. I can see how that might work. 

Harry: I called that Monday morning quarterbacking. He should have done that. Yeah. It would have been a better play, right. So here here's, here's one of my, one of my final questions is, uh, you know, what do you tell those people that say, tell me, tell me the first drug that ever got approved by AI.

Like, you know, how do you, how do you manage that? Because I get that all the time and I'm like, well, he wait for that. Like, it it'll be over.

Mark: Yeah. Yeah. I right. That's the thing. Their product development cycles are so long. If you. Wait for, if you, if your standard for, you know, accepting a new way of doing something as final FDA approval, um, you know, you're gonna be caught way behind the curve.

I think you have to evaluate the compounds as they've progressed every step of the way and say, You know, what are the results at this step? What are the results that that's done and are they moving forward? Uh, are they moving forward with better success or worst success or the same as the traditional process?

Um, I think that's, that's where we're at. I mean, that's one answer. The, you know, the other answer would be, you know, define what AI is. So, you know, there's. AI is sort of catch all thing. You know, some people would define AI as anything that's more complicated than you could do on an Abacus or something, you know, something like that.

Harry: So I, whenever I'm talking about AI, it's, it's a toolkit, right. And depending on what I'm trying to do, I may pull out a different screwdriver or a wrench or whatever, but it's a toolkit of, of. Capabilities, processes approaches that you can take to solve a particular problem.

Mark: Right. And people are applying in another areas of the, you know, drug industry looking, you know, um, using AI to develop biomarkers or are using AI, like patient recruitment things, you know, we're just applying it to, um, the first step and drug discovery.

Harry: So how do you think. You know, if you were a betting man, since you were a jazz working sort of on this stuff, you are somewhat of a betting man, because you're betting on something, finding something new is, is, you know, big pharma startups, where do, where do you, you know, I'm sort of betting on, on startups because I think they're much more nimble and quick, but, but you know, there are, there are, there are good papers in the space coming up from big pharma. It's just. No, I was trying to figure out where, where do you think the world is?

Mark: Yeah, I mean, there are certainly, you can fine exceptional to the robot. I think in general, most of the innovation is at small companies. Um, whether it's, you know, the small companies with the sort of the traditional ones that I mentioned, you know, some professors someplace had an idea that's being commercialized and then, you know, If it's successful, maybe they take it or maybe big pharma buys it.

That's why there's all these deals. And if you're looking at it, who's buying what, it's big companies buying the assets of little companies that did the, did the original innovation. 

Harry: But it's interesting though, because I think the buyer set has gotten has broadened. Right. Whereas normally I would think like it would be okay, Pfizer Merck, you know, like you go down the list right now.

I think there's, you know, potentially there's Amazon. Microsoft Google. I mean, I was just talking to someone and they were saying, yeah, we keep talking to Amazon about partnering, but you know, they're just missing data. And if they had data, like all of a sudden they become a competitor as opposed to a partner. So it's, it's a, it's an interesting, uh, dynamic of new names that are coming into the, onto the forefront.

Mark: Yeah, I think, um, you know, um, my experience, you know, was more with the pharmacy, but I don't, I don't doubt that at all, that, um, more and more people are getting interested and maybe, you know, maybe the AI.

Component of that is more accessible to people at Amazon and Google than it is to, you know, people in the top five pharma companies. Maybe they're just more, you know, familiar with, um, how to manipulate large, large amounts of data. 

Harry: I, yeah, but I, I, it's interesting because I think you have to have both, I mean, at some point it's producing all this data, but then someone needs to look at it.

And, and think about ok a yeah. Okay. That makes sense, right. Or, okay. I can believe that B how the hell are we going to test that? Right. What's the model and how are we going to, and then what's the rest of the process going forward. And that requires some, you know, I wouldn't want to novice. 

Mark:. That's fine. Well, I did try to help them with that, to put that piece, you know, the, the, the program or the platform came up with this and it was always what was helpful and what sort of distinguishes it is aside from producing, you know, I answer or quote unquote answer in terms of a prediction. You could trace back why you know, why the computer like this one or why the computer liked that one.

And so that was very helpful to me. And that helps you design the, the appropriate, um, test model to put the molecules in. But, but a lot of AI, they don't, it, you know, it doesn't give you the rationale. It's sort of a black box type of product, and that's much more difficult to deal with. 

Harry: So this must also spur a lot of IP generation. 

Mark: I mean, yeah, I, yeah, absolutely. I think it has the potential to do that. Like obviously I can't tell you. 

Harry: Yeah, no, no, no, no, no. I'm just saying in general, right. Uh, as you're moving down this road and you're identifying things at a, at a, a speed much faster than others, right. And the IP attorneys must be quite busy or you hope they would be, you would get that.

Mark: That would be a. Predicted outcome. Yes

Harry: Well, Mark, you know, any other thoughts along these lines that you can share with, you know, the people that are listening to this? I mean, there's, I asked physicians, listen to this. I have people in the pharmaceutical industry. I have my wife listens to it, right. As a lay person. It's, you know, anything you can share with that group from, uh, you know, Your experience that you, that you would want them to know?

Mark: I would just, you know, have an open mind and make your decisions like a scientist based on results. Uh, and if you do that, you know, you're, you're on a good path. 

Harry: Excellent. Excellent. Well, I really appreciate the time. It was great to talk to you and hear this, uh, this story. Um, you know, as these products go forward, um, you know, we may come back to you and ask you to be back on the show.

Mark: I'd love to come back. Excellent. Thank you. Thank you.

 

Episode Transcription

Harry: All right, Mark. Welcome to the show.

Mark: Thank you very much for asking me, huh? 

Harry: Yeah, no, it's great to have you here. Um, and I, and I honestly, I'm really excited about this conversation after Andrew mentioned, like, you know, he sort of gave me a sneak peek of, of, uh, The story. And I was like, we gotta do the show.

Mark: Like we gotta have this conversation. 

Harry: And so Mark, just, just for, you know, everybody that's listening, sort of, can you give us a little bit about, of your background because you're, you're not the, you know, computer science, uh, born and bred person, right? That a lot of the people that I will have on the show you're. You're hardcore drug discovery

Mark: I mean, I had one programming course in college that I actually took to fulfill a foreign language requirement because, you know, programming languages were the neat new thing and they wanted to encourage that. So yeah, I have no sort of technical knowledge in programming and my background, my training is in clinical pharmacology and, uh, basically spent the last 30 years in the drug industry in big pharma and CRO, uh, on.

Nonclinical preclinical development and clinical pharmacology PKPD. And, um, you know, before I started twos, I was consulting for a while. And before that I was at Jazz were 15 year or not 15 feels like, like 12 years, uh, starting out as VP of research and then had various titles along the way. And along the way I became inventor on something like that 30 patents, uh, including the ones, uh, the original ones for the anti-histamine, uh, Lycra, and then at, uh, uh, Jazz on some of their, uh, Xyrem patents and, um, just, uh, you know, the spent my time in developing drugs. That was my passion. And what I'd like to do, especially the early stages of drug development.

And then started, uh, consulting after I left Jazz and twoXar was one of my first clients. So that's how, sort of, how I got to, uh, got to them., 

Harry: So so essentially, you know what I would consider, you know, a traditional drug development, uh, background. Well, I mean, Obviously distinguished, like you've done some incredible stuff cause my son takes Allegro. So I'm sure he's very happy that you developed it. Um, but you know, I guess your first exposure to this was dealing with twoXar as a client. If I understand correctly.

Mark: Yeah, that's fine. So, you know, they were, um, mostly uh, computer and bioinformatics people. They had a few people with drug development background, but they needed, they realized they needed more in that area.

And through a friend of a friend, they reached out to me and I was, uh, consulting. And so I said, sure, I'll be happy to help them. And you know what I knew about AI at that time was like from documentaries. I had seen doc, you know, computers beating chess, and then there was a one you a more recent one about alpha go and about the masters and the team from the, you know, Google bought them and they've got they're playing the master and the, uh, computer program wins and they're oohing and nine about this and they can't figure out why the computer, you know, made this winning move or why it came up from there. Yeah. And you know, that was my understanding of AI and I couldn't understand how AI could possibly help with drug development, because I could see, okay, for a game, I can see how like a computer could play against itself over and over again, and figure out what, you know, what are the winning moves and learn to think ahead and things like that.

But for drug development, you know, if you use that game analogy and if you use like, um, success as an FDA approval, then there's only a few thousand games that have ever been played. And so how is the computer supposed to iterate on that? And if you define an approval and approved drug as a win as the target that you're going after, then how are you ever going to come up with a better drug or a drug for a disease where there, you know, there aren't drugs approved, which is the really the goal. So I, I didn't understand how AI could help with that. But as I became consulting with them, you know they obviously, they didn't bring me in for the AI part. They brought me in to say, okay, we find, you know, we found this interesting molecule and we have these, how should we test it? Or we've tested it.

And now we have these results, help us help us with that or help us with the next step. And, you know, I would do that and I would say, Oh, No, these, this looks pretty good. This is neat. I wouldn’t have though that, that might happen, but obviously it did. And then sort of retrospect, they can figure it out. Okay. So now we should do this, you know, and I would see that.

You know, that that pattern would repeat itself where they would, you know, they would run their computer platform and come up with like nine or 10 molecules and put them in a nonclinical model of efficacy and two or three of them would pop out positive and the, the hook for me was they all had new mechanisms of action.

So these were mechanisms of action that hadn't been tested in the patient population for that disease. And so, you know, that was an immediate challenge to my understanding of AI. So how did the computer come up with. That. And I didn't know the answer to that question, but you know, the more and more I saw results like this, the more I'm more, I thought, however, it's doing its producing results that are interesting, you know, at a, sort of a higher success rate than a lot of, uh, Uh, traditional methods of, of drug discovery. And that was something that was very appealing to me and I wanted to be a part of, so I started, uh, talking to Andrew then and, uh, sort of ended up as senior vice president of R &D here.

Harry: So what, you know, walk me through. Sort of you know that from skepticism to, you know, I want to be part of this it's, you know what you did, you said, look at you. We showing you pathways 

Mark: It was the pathways. It was the results from the past, but, but also my understanding of, of AI sort of changed. So, you know, I joined in, uh, November of last year, 2019, and in September of 2019, I was reading this article by Bob temple, who is, uh, this senior person at the FDA. And he's been around for a long time. He was involved with, uh, the approval of Allegra and he was talking about how there was this new age in drug development can call it the age of individualization or like personalized medicine.

Right? So recognizing that individual differences in patients, uh, contribute to individual variations in response. And they're there, you know, the, he called the previous ages, the ages of safety and efficacy, and he was citing regulation and he said, but for this age there was, it wasn't really kicked off by regulation.

It was kicked off by the discovery of drug interactions with Seldin and I was involved in doing those drug interaction studies. And it was actually those studies that led to the invention of Allegra, which didn't have the, you know, the problems that led Seldin to be with withdrawn from the market. But it, it struck me that what was happening then was like the very beginning of the introduction of signal detection into the pharmaceutical industry. You know, Seldin on average was very safe, uh, so safe that the company wanted to go OTC. And that's what I was hired in to do, to help them repeat some studies and get it ready for OTC. But then we got started getting these reports of drug interactions and, uh, arrhythmias.

And it was the first sort of application of signal detection in the pharmaceutical industry signal detection signals from individual patients. So looking at individual patients for four signals, and at first, you know, the, the signal detection system was so crude that the only signal that broke through were arrhythmias. And then we learned how to. Parse a piece of the ECG, the QT interval as a signal detection system. And, you know, as a result of those, uh, drug interactions, FDA also introduced MedWatch. Which is a signal detection system to get adverse event data from marketed drugs. And then, you know, that's one of the things, FDA monitors, how the drugs are actually working in the marketplace and are there any unexpected reactions and things like that.

And yeah, it's sort of hit me then that what twoXar was doing and what, how they were using AI was actually as a signal detection system, sifting through genomics data and phenotypic data and all these different datasets and examining it with an AI system to look for efficacy signals. And with that sort of reframing of you know, drug discovery as a problem in signal detection, it makes it a problem that's amenable to a computational solution and allows you to apply, you know, methods of signal detection that have been used in other places, other industries. Uh, to pharmacology and you know, the results that I, then it sort of clicked and it fit with the results that I was saying, Oh yeah, you know, the computer wasn't playing a game to come up with this compound.

It was sifting through this data looking for signals. And then, um, you know, we tested it and lo and behold, the signal was, was validated at least in the, the first, uh, animal model or in vitro test. And so with that sort of reframing of my understanding of AI along with sort of re sort of defining the drug discovery problem.

It started to all fit together. And that's when I thought, you know, wow. You know, if signal detection 20 years ago, 25 years ago, according to Bob Temple was the thing that ushered in this new era in, you know, development or pharmaceutical development. Well, if we apply signal detection to the beginning part to efficacy, you know, I think it just, it, it has tremendous potential and I'm just, uh, it's a price that hasn't been done before because it's really kind of, uh, the results are better than let's say just a much better than I would have anticipated going into this.

Harry Glorikian: Now, so it wasn't just necessarily the results, but I think if I heard you correctly, it's also this factor of, time being shorter

Mark Eller :. Yes. It was. It's much more efficient. I mean, you know, twoXar was a very small company and yet, you know, they've got 18 programs lunch and I think on 10 of them, uh, we have in vitro or, uh, animal pharmacology data with positive results and, you know, It would have taken years and years, especially for a little company, like, like this, to generate that data with traditional methods.

So it's like, much more efficient. It gets you out of the triangle that manager's always talking about, you know, cost and time and quality. And you can get any two of those, but you have to sacrifice on, on that third one. Well, this just sort of, you know, breaks that all open and you can get really good results, uh, are very fast and you know, much more efficiently than the, than the traditional approach. So it was, it was just a big jump, um, all the way around from the traditional way of doing things. 

Harry: Now, the other thing though, that you said is, and so I'm trying to use these two axes, right? One is time. One isyou are seeing patterns or pathways that you were like, ah, not seen that before and opening up sort of a unique area to look at that you might be able to develop a new molecule again, which of course is a always a great place to be

Mark: Yeah, exactly. So that was, that was the really the exciting part. So, you know, there's a lot of, a lot of drug development and a lot of drugs are like there, you know, me too, drugs, they're second generation or third generation of something that does this and maybe there's incremental improvements, you know?

And the third statin that it's approved as better than the first one or whatever, but they're working by the same mechanism. What was interesting to me is that, you know, twoXar’s approach and their platform would identify molecules that were potentially effective in a given disease. And the approved ones would show up to or stuff that might be in development by the companies would show up to, uh, which, you know, it gives you some comfort that you're on the right track.

But what also show up is stuff that not only that hadn't been tested, but, but that have mechanisms where there's no drug that's approved or even been tested in patients with that disease that have that mechanism. So it's a whole new way of sort of treating the disease. So it's like the computers come up with a new or the actions of the computer, the whole system.

Equal included have come up with a, like a new, a new hypothesis for what might work to treat a given disease that, that hasn't been tried before. And then we test those in animal models and it would. You know, test 10 of them in three of them pop up, uh, effective. That was very, very exciting. 

Harry: Yeah. I've, you know, I've talked to a lot of people where they have it, you know, they have a well understood process of making a molecule and then they, you know, to their system and say, well, you know, is there a faster way, more efficient way? And the system can sometimes tell you how to get to the same end result in a different way than anybody was classically trained.

Uh, that might bring down cost and decrease time and so forth. And so this doesn't sound, you know, I do can see similarities between all these approaches. Um, now the other thing though, that when I talk to Andrew, he's like, you know, his hypothesis was, you could shave off three to four years of time on the whole process. Like, do you agree with that sort of,

Mark: Yeah. I mean, from, uh, what we've done so far from, you know, the, the, uh, things that we've taken furthest along in and have results back from, uh, nonclinical pharmacology models. Uh, yeah. You know, it was like something like four months. From, you know, saying we want to investigate this disease to having results in a laboratory animals.

And, you know, doing them at CRO is the same. CRO is at big pharma uses in sort of standard classical models. And most of that time, frankly, was, uh, for the actual conduct of the study, not the, all the, um, stuff that. Brought us to that prediction and the stuff that, you know, You know, traditional process that I would bring you to, that prediction might take, might years.

And here, you know, we had collapsed all those initial steps down into one that just generates efficacy predictions that can be, you know, immediately tested. So, so yeah, there's that sort of time and efficiency aspect of it. And then, you know, when there are new molecules or new mechanisms of action that made it really, really exciting.

Harry: So now, I mean, but at some point the system generates what it suggests. And then the human being though has got to look at it and be like, 

Mark: Yes of course. Yeah. Okay. So, you know, I can give you like a flow step and the first step is like all the computers. And then the last step is basically, you know, humans deciding what animal models should we use to test this hypothesis. And then the intermediate steps, you know, span the range of more computer and less human to more and more human. So, yeah, uh, after the initial, you know, generation of possibilities, which might be, you know, A thousand come out or something like that. Then there is this winnowing process, uh, and ranking process.

And the first few steps of that are also computer assisted or AI assisted to rank them or, uh, eliminate molecules that might be too toxic or for a given indication, things like that. And there there's human intervention along the way until finally the decision as to which of the molecules to put into a non-clinical model is, uh, dependent on human insight and, uh, not, not AI.

Harry: So, but at some point that information now you are working with the people actually building the models, right? So there's yes. I assume that, that, you know, I think of this as a figure eight, right. It just, at some point there's feedback, there's correction or there's modification of the, and then it just keeps going back and forth then just makes it better.

Mark: Yes, exactly. 

Harry: I mean, it was interesting cause I was talking to someone earlier today from a Big pharma and saying, you know, one of the companies I was talking to has said, they're constantly improving their, their algorithm. And he says, nah we don't do that at Big pharma because, you know, our model is pretty well set. And, and, you know, we're, unless we think that there's going to be some huge breakthrough. 

Mark: I know, I think we're on, I don't know. This is a question for Andrew, but version. I don't know, but I I'm sure. Pretty sure it's in, you know, hundreds of iterations of the, um, stop where, yeah. So there's this constant learning process.

Harry: So do you, do you see this actually also affecting cost in development? 

Mark: Well, I think, you know, cost and timing sort of go hand in hand. It really, the way you achieve the tiny is because you're eliminating all these, you know, wet lab rate, limiting experiments, and those wet lab rate-limiting experiments also cost money. So yeah, there's, there's a big, um, cost savings as well as, as well as a time saver. 

Harry: But yeah, I think about it from a, uh, uh, uh, rate limiting, but also. I think the number of parameters, these algorithms can look at it as much more than, you know, definitely, definitely more than me. Right. 

Mark Eller: I mean, that is how it comes up with new stuff that you and I haven't or other people haven't thought of.

It's, you know, I think, um, for one of the, uh, programs, um, lupus, I think, uh, they were like, Uh, Tom was talking to Erin who already, who heads up there, runs the platform. And it was something like 2.5 billion pieces of data that, uh, were going into this, that it was sifting through. So yeah, it's, it's more than more than humans can handle.

Harry Glorikian: So, but, and then I always think about like, the papers that have to be coming out of this. I mean, at some point, you know, you got to start to let the world know that there's this other potential pathway that you could use this, you know, or, or, you know, just to publish this stuff and say, here's a different way to come at these problems and make it more of a widespread now I know, you know, you know, as a startup, you want to be first and own it all. But I think about that just from a science perspective.

Mark: Oh yeah. Yeah. I agree. I mean, there is, you know, there's that tension between proprietary and getting the information out. And so in terms of results, that's certainly something, uh, that we are publishing, for example, the lupus one on, I know that the abstracts have been presented or, you know, all the meetings are virtual now.

Uh, yeah. Uh, yeah, that where we've presented the, data, uh, from those studies and also a couple of others, like, uh, cellular caution, Noma. And so the data is starting to come out and people can, you know, Judge for themselves and take a look at it's very positive. 

Harry: So now, uh, for, if you were talking to other people like with that, that have your background, so what, you know, how, how do you, and I'm sure it's come up in, in you've I'm sure you have this conversation with, with, uh, colleagues.

Yeah. It's what do you say to them and how do you frame it in a way that they sort of can get there? Their head around it, um, quickly, if right. And, and, and what are the skeptical comments that you get? Cause I'm sure that you and I get probably the exact same comments. 

Mark: Yeah, I think it depends upon, you know, it depends, it depends upon what they're looking at. So if you're, if they're looking at the results of a particular program, if they're looking at our SLD data, I just say, you know, just, just look at the data and decide for yourself and then ask yourself, does it really matter? You know, if I came up with the idea or a computer came up with it, so there was this, this game that I used to.

Um, play at Jazz. We were doing like, uh, opportunity assessment and we called it doing a pre-mortem. So we pretend, okay. We, we bought this asset and we did everything and executed according to plan and it failed. Why did it, why did it fail? And if you write down all the reasons it could potentially fail.

And you know, the last reason is because the computer came up with the idea. Yeah. That wouldn't stop us from moving forward with that program because in the end, you know, the FDA doesn't care. If a computer came up with the idea or, you know, whoever it, they care about a development plan about, um, development, rationale and your data.

So I, it's very easy to convince people on individual products. I think on. You know, the AI side, it's, it is more people are more skeptical and they're more like in my position and I have to say, look it, you know, put away your preconceptions about what AI is and you know, if you're like me and you thought it was about computers, playing games against humans and they, you know, it's a little bit, you know, maybe that's part of it, but it's a little bit more than that.

And if you view it as signal detection, Um, that might help you understand what the, what the potential is. And, you know, for me, that helped and also seeing the results, you know, repeat over and over, you know, convinced me that that was, uh, that was in fact the case. But yeah, we do run into people who are skeptical and say, Oh yeah, you know, this data looks really nice. And that piece of data looks really nice and this program looks good, but you know, they're not. I guess convinced that there's a block there and they can't see that, uh, you know, if it works three times or five times or 10 times that it's likely to, you know, keep working and, you know, we think that there are, you know, a thousand diseases that it might be applicable for.

There's some probably that it's not, you know, like infectious diseases, it's the platform. It's not set up to do something like that because you've got a third thing, the organism itself. But, uh, for other things, it's yeah, it's very broadly applicable. 

Harry: Yeah. I mean, uh, you know, we've always talked about it. Identifying new targets. Completely different pathways that, you know, have a major effect on the disease that nobody ever considered. Right. Um, and then, uh, the third is, is repurposing something that's already out there that might actually have a meaningful effect on this particular disease, but nobody is nobody's really using it in that way.

Mark: Right. Yeah. Yeah. You know, sometimes we come up with stuff like that too, you know, we'll be in our list of things that are predicted to be potentially effective. There might be, you know, a drug that's, um, on the market for something else. So that does happen occasionally. 

Harry: So, so, if you were coaching somebody and say, and that was working their way through the system. Would you tell them to study something? Would you tell them to. I don't know, read a particular book. Would you, how would you sort of coach someone along that was the you're younger? You? 

Mark: Um, I would say, always be guided by science and, you know, look at the results and look at the hypothesis and connect the dots. And if, you know, if you connect the dots and it says signal the texting AI works, you know, trust that and believe it and put away your preconceptions. You don't have to understand the details of the programming to see the results. And if you can just understand the inputs, and if you can understand that the idea that the computer can take two and a half billion pieces of data and process it, and you can't do that, um, And comes up with predictions.

Well, test the predictions. It's sort of empirical it's, shouldn't involve, you know, I'm a scientist, it doesn't involve belief or trust or whatever. It's like, here's the, here's the output. Did the output work? Yes, here's the input and this experiment here's out, but did it work? Yes. How many times do you need to, to see that to, to be convinced and, you know, you can just sort of be agnostic in your, uh, beliefs about AI or computers or traditional approaches and, and just, you know, just because you're used to doing something one way, And it's worked, although maybe, you know, slow and inefficient, you know, don't be closed to the idea that there might be other ways to do things that, uh, get you out of that triangle of cost and time and quality.

Harry: Yeah. Another thing that I think about is the systems are it's, it's not like they're standing still either. They're in a constant state of improvement and evolution, which is just making them better over time. And we are collecting more data sets that the system can ingest and build into the model. So I think it's just, it's moving forward at a, at a pretty, at a very fast pace. I mean, you know, I try, I try to explain the, the speed at which things are happening to people. And it's very difficult for the human mind to understand. Doubling doubling, like has a ha I don't know why we have a hard time getting around.

Mark: I mean, it's like, you know, science, sometimes things in science moves sort of in a linear, predictable pattern. And sometimes it's just, it jumps like, you know, the, cell vein drug interactions leading to the MedWatch system and signal detection for safety. What that, that didn't come up out through a linear process.

That was a jump and a response to a problem. And the solution sort of developed Denovo. And this is sort of, I think, a similar situation, it represents sort of a break from the traditional way of doing things. And you, you know, if you're open to an objective to looking at results, you should be sort of.

Okay with that. Um, you know, but there, you know, we get comments from, you know, you know, where's your, you know, five KOL’s in this disease area who came up with this thing, you know, the whole purpose of this, you know, we're a small company and we've got, we've got stuff in. That we're developing for SLE. We've got it for oncology indications. We've got some, you know, you can partnered with, uh, people in, in various therapeutic areas. We don't have, you know, we have, we hire a KOL as consultants and we need them, but we don't have this, you know, staff of therapeutic disease experts who have. You know, working for 20 years to come up with this molecule, but you know, there there's, you know, to give people credit.

There's a lot of companies in the Bay area that that's how they are start to some professor from Stanford and working on this problem for 20 years. And then they spin it off into a little startup company. And people can, you know, wrap their head around that, whereas, okay. Uh, Andrew and Aaron were working on this computer system and now we're going to apply it to pharmacology and we can apply it to oncology.

We can imply it, uh, inflammatory diseases and we can apply it in other areas and, you know, we've got results and that just, just it's different. 

Harry: Yeah, yeah, yeah. Yeah. I mean, I think in the other part is, is. If some new pathway comes up, that the system came up with. I mean, I'm, I'm hard pressed to find a KOL that really like, you know, would have gone. Oh, Oh yeah. That was, we absolutely knew about that. Right. It's what I find is they're like, Oh, I mean, I didn't even think about that one. And yeah, now that you're showing me the data that's short of, I can see how that might make sense. It just wasn't. Right in front of them. 

Mark: Yeah. It's what I would call retrospective predictability.

Now I can't think of it, but now, now that now that the computer did and all that, you've got positive results. I can see how that might work. 

Harry: I called that Monday morning quarterbacking. He should have done that. Yeah. It would have been a better play, right. So here here's, here's one of my, one of my final questions is, uh, you know, what do you tell those people that say, tell me, tell me the first drug that ever got approved by AI.

Like, you know, how do you, how do you manage that? Because I get that all the time and I'm like, well, if wait for that. Like, it it'll be over.

Mark: Yeah. Yeah. I right. That's the thing. Their product development cycles are so long. If you. Wait for, if you, if your standard for, you know, accepting a new way of doing something as final FDA approval, um, you know, you're gonna be caught way behind the curve.

I think you have to evaluate the compounds as they've progressed every step of the way and say, You know, what are the results at this step? What are the results that that's done and are they moving forward? Uh, are they moving forward with better success or worst success or the same as the traditional process?

Um, I think that's, that's where we're at. I mean, that's one answer. The, you know, the other answer would be, you know, define what AI is. So, you know, there's. AI is sort of catch all thing. You know, some people would define AI as anything that's more complicated than you could do on an Abacus or something, you know, something like that.

Harry: So I, whenever I'm talking about AI, it's, it's a toolkit, right. And depending on what I'm trying to do, I may pull out a different screwdriver or a wrench or whatever, but it's a toolkit of, of. Capabilities, processes approaches that you can take to solve a particular problem.

Mark: Right. And people are applying in another areas of the, you know, drug industry looking, you know, um, using AI to develop biomarkers or are using AI, like patient recruitment things, you know, we're just applying it to, um, the first step and drug discovery.

Harry: So how do you think. You know, if you were a betting man, since you were a Jazz working sort of on this stuff, you are somewhat of a betting man, because you're betting on something, finding something new is, is, you know, big pharma startups, where do, where do you, you know, I'm sort of betting on, on startups because I think they're much more nimble and quick, but, but you know, there are, there are, there are good papers in the space coming up from big pharma. It's just. No, I was trying to figure out where, where do you think the world is?

Mark: Yeah, I mean, there are certainly, you can fine exceptional to the robot. I think in general, most of the innovation is at small companies. Um, whether it's, you know, the small companies with the sort of the traditional ones that I mentioned, you know, some professors someplace had an idea that's being commercialized and then, you know, If it's successful, maybe they take it or maybe big pharma buys it.

That's why there's all these deals. And if you're looking at it, who's buying what, it's big companies buying the assets of little companies that did the, did the original innovation. 

Harry: But it's interesting though, because I think the buyer set has gotten has broadened. Right. Whereas normally I would think like it would be okay, Pfizer, Merck, you know, like you go down the list right now.

I think there's, you know, potentially there's Amazon. Microsoft Google. I mean, I was just talking to someone and they were saying, yeah, we keep talking to Amazon about partnering, but you know, they're just missing data. And if they had data, like all of a sudden they become a competitor as opposed to a partner. So it's, it's a, it's an interesting, uh, dynamic of new names that are coming into the, onto the forefront.

Mark: Yeah, I think, um, you know, um, my experience, you know, was more with the pharmacy, but I don't, I don't doubt that at all, that, um, more and more people are getting interested and maybe, you know, maybe the AI component of that is more accessible to people at Amazon and Google than it is to, you know, people in the top five pharma companies. Maybe they're just more, you know, familiar with, um, how to manipulate large, large amounts of data. 

Harry: I, yeah,  but I, I, it's interesting because I think you have to have both, I mean, at some point it's producing all this data, but then someone needs to look at it.

And, and think about ok a yeah. Okay. That makes sense, right. Or, okay. I can believe that B how the hell are we going to test that? Right. What's the model and how are we going to, and then what's the rest of the process going forward. And that requires some, you know, I wouldn't want to novice. 

Mark:. That's fine. Well, I did try to help them with that, to put that piece, you know, the, the, the program or the platform came up with this and it was always what was helpful and what sort of distinguishes it is aside from producing, you know, I answer or quote unquote answer in terms of a prediction. You could trace back why you know, why the computer like this one or why the computer liked that one.

And so that was very helpful to me. And that helps you design the, the appropriate, um, test model to put the molecules in. But, but a lot of AI, they don't, it, you know, it doesn't give you the rationale. It's sort of a black box type of product, and that's much more difficult to deal with. 

Harry: So this must also spur a lot of IP generation. 

Mark: I mean, yeah, I, yeah, absolutely. I think it has the potential to do that. Like obviously I can't tell you. 

Harry: Yeah, no, no, no, no, no. I'm just saying in general, right. Uh, as you're moving down this road and you're identifying things at a, at a, a speed much faster than others, right. And the IP attorneys must be quite busy or you hope they would be, you would get that.

Mark: That would be a. Predicted outcome. Yes

Harry: Well, Mark, you know, any other thoughts along these lines that you can share with, you know, the people that are listening to this? I mean, there's, I asked physicians, listen to this. I have people in the pharmaceutical industry. I have my wife listens to it, right. As a lay person. It's, you know, anything you can share with that group from, uh, you know, Your experience that you, that you would want them to know?

Mark: I would just, you know, have an open mind and make your decisions like a scientist based on results. Uh, and if you do that, you know, you're, you're on a good path. 

Harry: Excellent. Excellent. Well, I really appreciate the time. It was great to talk to you and hear this, uh, this story. Um, you know, as these products go forward, um, you know, we may come back to you and ask you to be back on the show.

Mark: I'd love to come back. Excellent. Thank you. Thank you.