Faster, Cheaper Drugs with AI

to it. But when you zoom out, it's strange because, as I've said before on this show, and as I will say again on this show, one of the main things technology does is it makes things more efficient and therefore cheaper. Over the past few centuries, we've seen technologies make all kinds of things cheaper, everything from clothes to food to TVs. So why hasn't new technology made drugs cheaper? Two.

I'm Jacob Goldstein and this is What's Your Problem, the show where I talk to people who are trying to make technological progress. My guest today is Alice Zang, co founder and CEO of verge Genomics. Alice's problem is this, how do you use artificial intelligence to drive down the price of discovering and developing new drugs? Why is it getting more expensive to develop drugs, despite the fact that

we have better technology to do it. Yeah. Absolutely. One of the reasons is, you know, even though a lot of the new technologies we've developed have made us better at testing more drugs faster, but the fundamental problem is that even if we can get a drug all the way to clinical trials, which is the last step of drug development, ninety percent of those drugs still fail. So if you think about it, we're spending millions on each drug. Of those drugs are failing at the last and most

expensive stage of drug development. And so really most of that billion plus dollar figure you hear is due to the cost of failure. Just to be clear, that figure more than a billion dollars. It's you've got to include the cost of all the drugs that don't work exactly, the ones that do right exactly. So the ones that do work have to pay for all the ones that fail. That's the fundamental problem, exactly, And you're setting out to

fix that if you can. Absolutely, we think there's an opportunity for AI to fundamentally shift really the failure rate, and the most impactful time to do that really is the failure in clinical trials. So can we predict before we go in to these expensive clinical trials genes or targets or drugs that are more likely to work in humans, because even a ten percent decrease in that failure rate could have massive I saw a number of up to

fifteen billion dollars annually in industry cost savings. You could still be in a universe where most of the drugs that go into clinical trials fail, but instead of ninety percent of them failing, seventy percent of them fail, and that would be a huge win. That would be a

huge efficiency gain. It would save a ton of money, absolutely, And I think that's something that's underappreciated about AI and really any technology, is that oftentimes people have this expectation that this technology is going to absolutely transform a field overnight. And I think what people don't appreciate is that most of the time that doesn't happen. It's always step by step incremental. But even a ten percent change would have billions of dollars of cost savings and would be a

huge win for patients in the industry worldwide. I like that frame, actually, I like that frame of maybe AI can have drugs fail most of the time, but not as much of the time as they fail. Now, like, it seems very credible, It seems very plausible. Would you put it that way? Yeah, it's all life is nothing but a learning process, Yes, getting less bad at everything. So I know you were studying to be a doctor and a researcher not that long ago, a few years

ago before you started your company. Like, tell me how you went from an mdphd program to starting the company. Well, my PhD research was actually in using genomic analysis and computational biology to analyze large scale data sets and find

new drugs that could improve drug development. And we found that from our very first drug that was predicted from our algorithms when we put it in mice after they've been injured, help them walk and recover from that injury, that nerve injury about four times faster than the leading standard. And that was just the first drug that was predicted. And I looked at this technology in this approach and

I thought, Wow, there's so much promise here. You know, am I really going to just publish this and let it sit on a bookshelf somewhere, or if I'm not going to be the one to really develop this to patients, you know who will, And when I looked out off the field, I did not see a ton of biotech or farmer companies that were truly computationally driven. Usually within pharma companies they might bring in computational biologists to support.

There are scientists or their biologists, but there wasn't really a genomics computationally driven company at that time. Now there are many, but at the time there are very few. And so I actually, you know, it wasn't a binary decision. People always ask me, how did you make the courageous decision to leap? It wasn't really like that. I think what we did first is that we just took three months three month leave of absence. We joined a program, an incubator called a y combinator. We as you and

you and well me and my co founder Jason. And the first question really was, you know, can we even generate some data that validates that computational biology can predict targets that work? And then when we saw some data, the next question was can we even hire people that want to come on? And the next question was can we even raise money from people that will care? And I think that is so such an important lesson because I think people oftentimes get caught up in just the destination,

you know, is where I want to be? Is this the career I want to have that they don't take the first step, And really it's the first step that's needed to actually get the data to even decide if it's the appropriate track for you. And did you really just keep thinking, well, this might not work, but let's do the next thing. Were you in a place where you could have gone back to the MD PhD program

for a while. Yeah. I took a leave of a continuous leave of absence for probably over five years, probably more than I should have, until the point where a lot of my friends are like, are you really, are you really gonna go back? And finally the medical school is like, you're not really going to come back, let's

just terminate your leave of absence. But it was in the first few years a really important safety net for me that gave me the psychological safety to really take a risk and really pursue a new idea that I don't know if I would have otherwise. And I think that's so important. I think for universities to provide is that to recognize there can be more than one track for people to do really excellent science and make an

impact more than just becoming a professor. And sometimes that psychological safety is what's needed to help people find their ultimate calling too. By the ways, so far, By the way, what's a very brief definition of computational biology. It's really, at the end of the day, in my view, just the use of computers and data sets to understand and biology better. By the way, what happened to that molecule that you were testing in mice in grad school? That

seemed useful? I don't know. It's a good question. Actually, I think the project was taken on by someone else, but I'm not actually completely sure. So, Okay, you leave grad school, you start a company you in fact now have taken You have a bunch of molecules that you're working on, and that seemed promising. But there's one that is in clinical trials now right to treat als Luke Gary's disease, a terrible disease that is very poorly treated.

And I thought that we could talk about the story of that molecule of that drug as a way to understand the way your company works. Can you just sort of take me through the life of that drug? So far? Yeah, absolutely. I'll start off just by talking about als and why it's been so hard to discover the right therapy, and then you know why how we did that differently. So, as you might know, LS Luke Garrig's disease is a

really horrible disease. What happens is that these neurons called motor neurons start dying, and most patients experience paralysis and then death, usually within three to five years of diagnosis. A very fast progressing disease, and there really aren't any meaningfully effective treatments that really slow or stop the disease today.

So a very horrible disease with a horrible prognosis and no available treatments, and why it's been so hard I think to discover really effective treatments is really just the complexity of the disease, and really any disease of the brain, the brain is the most complex organ in the body. So you end up having a lot of drugs brought into clinical trials that worked in mice. I always like to say we've cured LS or can There are many diseases in mice a thousand times, but none of them

have really worked in humans. So what we did differently was we started from day one by collecting data from over a thousand ALS patients as well as controls, and specifically, we collected samples of brain tissue as well as spinal cords from these patients that actually passed away from ALS. So you got samples from a thousand patients who had

died of ALS. How did you do that? So what we've done over the last seven years is we've signed partnerships with over twenty one different brain banks, hospitals, labs, academic centers worldwide that collect these brain tissues. They're usually donated from patients that have passed away from the disease and whose families want to contribute to research. Could So step one basically is get tissue samples from real patients.

And you said controls as well, right, So tissue samples from healthy people as well, so that you can use them as a basis of comparison. You have the samples, Now, what's step two? So step two is that we put an enormous amount of effort into quality controlling these, So that's a big underappreciated step. They can be very noisy samples.

And then step three is that we sequence them, so we profile, what is the expression of all twenty thousand genes in the genome, and we also sometimes do DNA sequencing, we look at genetic mutations. We also have a clinical information about that patient, how long did they have the disease, when did they die? And that makes for a very rich, multidimensional data set, and that gives us essentially a global

snapshot of what happened in that patient. H okay, and you and presumably the sequencing that you're doing on the patient's tissue samples, you're doing the same sequencing on the controls, the samples from healthy people. So now you have this very large data set. What's the next step. So then you have this snapshot of what happened, and the tricky

part is to figure out what caused it. I often liken it to a plane has crashed, right, You're looking through the rubble and you want to figure out how the plane crashed and how that information can be used

to prevent further planes from crashing. So that's when our software engineers and data scientists as well as machine learning scientists come in and we have algorithms essentially to integrate multiple data types all the way from the RNA, so how the genes were expressed to genetic mutations to essentially create a map of disease biology, and within the map our networks of genes that are all interconnected that we

believe cause disease. And so I like to think about it like when you're looking through a plane crash the rubble, you want to find the black box, which I'll help you figure out the cause of the disease. And by having all the information, we essentially locate the black boxes of disease, the targets that are really at the center of those networks, and then we design drugs against those

targets that we believe can reverse disease. It seems like differentiating between correlation and causality in this particular setting would be really hard, right, Like to use the plane metaphor, if you had a bunch of planes that crash and a bunch that hadn't crashed, you might say, oh, like the wings were off all the ones that crashed, and that's why they crashed. But actually the wings came off because they crashed, right, and it was something else that

caused the crash. I feel like that would be I mean, an obvious problem. Yeah, that might be hard. To solve Absolutely, you hit the nail on the head, and actually the plane metaphor is a really great one here. For one of the biggest challenges with looking at tissue from a patient that already died is that you're getting the crash right. You're not seeing video of before the crash. You're really

getting the crash. And the challenge is how do you figure what caused the crash versus as well was just the effect of the crash, like a burned wing, etc. And one of the ways we do that is we combine different data types. So we found that looking at one type of data, for example, just RNA data is in particularly helpful, but it's actually looking at where do you get convergence signal that pulls through multiple types of data to start revealing more compelling signal. So as an example,

we look at genetic data as well. So genetic data is useful for looking at cause versus effect because it contains information about genetic mutations that you were born with as a baby that then lead to increased risk later in life for a disease. And that's kind of nature's human experiment for really cause and effect. And when we layer that on that information on with the RNA data.

It actually gives us information about how the genetic drivers are acting in these functional pathways, which is a big issue actually with just looking at genetic data on its own. So I wish I had a better I wish I had a way to actually string that through to the plane metaphor. But and there's a time for leaving metaphors behind. Your company uses AI in drug discovery. I appreciate in a certain way that you haven't said AI yet, but

also I don't want to not talk about it. I mean in the sort of figuring out what's going on in this step? Is that well, is that the first instance in this process where you're using AI? Is it? We're talking about that here? Yeah, I mean, I think AI is a really broad term for any kind of

process where the computer is learning from something. So there are all sorts of applications of AI in this entire process, for example, how we're integrating the data sets together, how we're inferring what are the central nodes or the key targets. I would say the most classical use of A on the way that most people think of it is then once we have this network of say one hundred genes,

how do we actually find what the cause is? How do we find what is the hub or the right target to hit to turn off or on all hundred of those genes. And that's where machine learning and AI comes in handy. In a minute, Alice explains how this actually works in the case of the ALS drug verges working on. Now now back to the show. So okay, Alice and her colleagues at Verge have collected all these

tissue samples from ALS patients. They've used the samples to generate this huge data set that shows genetic variation and changes in how genes are expressed, along with lots of clinical data about the patients, and then they build these basically these AI models to try to figure out where in this complicated biological process that's happening in this disease, where they should try to intervene with a drug, Basically

where they should try and target a drug. I think of this oftentimes, like if you think of a map of all the airports in the US, you want to figure out how to go after the hubs like Chicago or New York. You don't want to go an airport in Kansas or I will wouldn't be very effective at

stopping airplane travel in the country. So there's a lot of different pieces of information that we collect to then infer what are the best genes that are not only central within this network, but also there's independent evidence of a disease causal effect or a relationship to disease. And so you do all that in this instance, and what

do you figure out? So what the algorithms spit out is essentially a ranked list of targets, all right, So these are ranked list of targets that are predicted if we could dry them, would restore that network back to levels of healthy people and potentially slow or stop the disease.

And then what we do is we take those targets and we start testing them in the lab, all right, So we actually what is kind of cool about the platform is we get all these targets from human brain tissue and we also can test them in human brain cells in the lab. So you get a list it's basically genes to target. You either it says upregulate or make this gene express more or make this gene express less.

Is that basically what the AI is out putting exactly, Like, so how long in the instance of this ALS drug. How long was the list? More or less, our initial set of targets was twenty two high confidence targets, and then we actually then generated another chut choosing updated data

of about thirty more targets as well. And what was really striking when we tested these targets is that when we tested them in the lab, we found that on average over sixty percent of them, though more recently actually around eighty percent of them actually validated in the lab, so they actually protected ALS patient cells from dying, which is very high. So we're really excited that we're actually seeing very robust validation of the computational predictions, at least

in the lab. Okay, so you have this list, you're testing it, something like half of them seem promising, you said, sixty percent seem promising. What happens next? Okay, So what happens next is that we so we test them in

these human brain cells. We understand the mechanism. One of the really interesting findings from this ALS program and specific is that when we looked at the network that we found in these patient spinal cords, we found a new cause of disease that was previously unknown so most of the hypotheses in ALS, where many of them to date, have really been focused around these protein aggregates, these clumps of proteins that we can easily observe by ie that you see in ALS patients. Right, A lot of them

are observational hypotheses. But what we found by looking at a deeper cut of the data is actually, at baseline, most of these patients actually had a baseline dysfunction in their life csomal pathway, which I like to call the garbage disposal pathway. It's what is critical to clear out junk from the cell. And because patients were at baseline vulnerable to these toxic insults, it wasn't so much the

protein clumps that were directly causing it. It was because they're already vulnerable to these clumps of proteins that their cells started dying. And is the idea that the gene you're targeting is causing the cell's garbage disposal to not work, right, Like you're trying to fix the garbage disposal by targeting this particular gene. Yeah, it's a central regulator of that pathway. And it was also a target that was ranked I think it was ranked number one or number two on

the list. So just to be clear, how how do you get from you know, so you have fifty or so things to test, fifty or so targets, something like thirty of them seem promising. How do you decide which of those thirty to proceed with? Yeah, so that's a great question. We get asked that a lot. I think at that point it's a strategic decision. Right, you were a startup, Right, we have to be able to develop

things quickly and capital efficiently. So we were lucky in that sense that one of the top targets was also a target that already had where the path to developing a drug was relatively smooth, A lot was known about that target. We could start doing chemistry and designing molecules relatively easily, and the target itself had actually been tested in the clinic for other diseases, not als, but things like Crohn's disease and surrounds, so we did know there

was some safety data around hitting that target. We do then for targets where we can't develop all of the targets, right, we can only take focused bets for targets where there's a bit more technical risk, Right, It might be a bit more exotic. People don't really understand how it works. There's not a lot of tools out there to really develop drugs against it. That's where we might partner with

a pharma company to develop those targets. And we have such a collaboration with Eli Lily where we developed our als target, but actually Lily has the opportunity to essentially take you targets number three through twenty two plus and choose four of them to develop themselves. Oh interesting. So in that way, you're essentially laying off the risk to this giant pharma company that can afford to make more bets.

I'd say we're distributing the risk and we're allowing us to really capitalize on the entire opportunity all of the targets, because it's impossible for any small startup to do, you know, thirty different programs. And it's actually in line with what a lot of pharma companies are looking for. A lot

of pharma companies are looking for. What is that novel target that no one else is working on that's kind of unexpected, Where if we could really get a competitive edge in here, this would be really meaningful for a position within within drug development in the next ten years. Well, and I mean it also seems compelling because even though this seems like a more promising way to do drug development, drug development is hard enough that anyone candidate drug is

probably not going to work, right. Yeah, An any biotechniqus to be able to have a pipeline and the ability to withstand I think some failures because I think it's unrealistic to expect one hundred percent of what you try will work. But that doesn't reflect on the technology itself, and that can be something unfortunate in biotech, where you know,

if the first thing fails, everyone's all can be. It can be tempted to say, oh, the technology didn't work, but in reality, you think about how many different drugs that pharmac companies test all the time. Right, So I think really promising technologies need to be afforded that runway and that ability to really take multiple shots on goal before you can get the end to really see if

it's working. Right. Well, I mean, if nine of traditionally developed drugs fail once they get to clinical trials, you could be way better but still likely to fail on anyone drug. Yeah, Yeah, even a fifty percent would be huge, right, but still that means one out of two drugs will fail. Relative to the world we live in now, a world where one out of two drugs fail could be a world where we get more new drugs for less money.

In a minute, the Lightning Round including the worst thing out being named to the Forbes thirty Under thirty, and the best thing about accepting that your company might sail. That's the end of the ads. Now we're going back to the show. Let's let's close with the Lightning Round. You personally interviewed over a thousand people when you were starting your company, as I understand it, which seems very intense. And I'm sure as if there's anything in your life

outside of work where you've been that intense. Oh, everything that is a core to my being. If you ask my spouse, you would say any new game that we start playing. And I'm very competitive and it's just part of my being. I iterate, I get a lot of reps in He always likes to make fun of me that I have an AI in my head. I'm constantly learning and improving the model until eventually I become a

lean mean. We've been saying a lot of Katan recently, and I think if we him fifteen times in a row, So yeah, I am very intense and thorough in my life. Is chat GPT overrated or underrated? Both? Actually? I think it's both over and underrated. It's overrated for some applications and underrated for others. I think it's overrated for things where there aren't a lot of information available already on

that thing. I think it's underrated for applications at coding, where there's already a large body of literature out there. So it's really good at replicating things that exist, less good at discovering new things that don't exist. I read an interview where you said one of the things you've learned as in running your company is you learn to be okay with your company dying with your company not making it, which I found like very surprising and interesting.

Can you just tell me a little bit about that. Yeah, I mean, I think it gets to really the core of how we drive our culture, which is I think that soul for so long companies have been driven through fear and bravado of you know, we're crushing it, We're pounding on our talking about how we're crushing it, and less about emotional vulnerably and introspection and self awareness, and ultimately I found the thing that really transformed my leadership style was learning what I had grips over of where

I was really attached to outcomes, And ultimately, I think for all CEOs, a lot of that is tying meaning to what happens with the company. If the company fails, this means something about me as a person, and I think that stifles a ton of innovation and curiosity and

tends to drive those cultures of fear. So ultimately, the thing, for example, that got me to stop micromanaging was really being okay with the company dying, because ultimately, what is micromanaging if not just fear right or fear or control. And once you let go of that fear and you

recognize you're just open to learning. You can still really want the company to succeed, and you can be passionate about it, but you're no longer thinking, oh, I'm screwed, or like I'm a failure if this fails, and that just opens a whole new level of levity and lightness. Nice. What's the worst thing about being named to the Forbes thirty Under thirty list? I think they did a photo shoot where there was a there was a very revealing split on the dress, and I still get constantly made

fun of by my close friends for that. What's one example of a thing that went wrong as you were building the company? Something bad that happened? Oh so many things. We had a whole period where there was a ton of attrition and people leaving, and you know, the first time that happens to a founder can I took it personally, It's like someone leaving your baby, and you wonder why. That was actually a huge growth moment for me because I was for so long trying to put for the

strong face. If it's okay, it's okay. And finally, at the end of like a month of this, I just sat in front of the company at an all hands and I honestly I just broke down in tears. I said, I feel like I failed you guys. You know I'm still grieving this. I really don't know what to do. And it was paradoxically in that moment, most of the team really rose up to the occasion and I found support in ways I didn't even know where possible from the team. Alice saying, is the CEO and co founder

of verge Genomics. Today's show was produced by Edith Russolo. It was edited by Sarah Nix and Lydia Geancott and engineered by Amanda ka Wong. You're always looking for more guests for the show. If there's someone out there working on an interesting technical problem with big stakes, tell us about that person. You can email us at problem at Pushkin dot fm, or you can find me on Twitter

at Jacob Goldstein. I'm Jacob Goldstein and we'll be back next week with another episode of What's Your Problem.