Making Undruggable Targets a Thing of the Past

Daniel Levine (00:00.398)

The Life Sciences DNA podcast is sponsored by Agilisium Labs, a collaborative space where Agilisium works with its clients ranging from early stage biotechs to pharmaceutical giants to co -develop and incubate POCs, products and solutions that improve patient outcomes and accelerate the development of therapies to the market. To learn how Agilisium Labs can use the power of its generative AI for life sciences analytics,

To help you turn your visionary ideas into realities, visit them at labs .agilicium .com.

You're tuned to Life Sciences DNA with Dr. Umar Drobid.

Daniel Levine (01:08.718)

Amar, we've got Jen Nwankwo on the show today. For viewers not familiar with her, who is she? Jen is the founder and CEO of 1910 Genetics. Before founding 1910, she worked at the intersection of life science, technology, and business. Her prior roles include management consultant at Bain & Company, director of business development at a health tech startup, and drug discovery stints at Novartis and Eli Lilly.

She has a PhD in pharmacology and experimental therapeutics from Tufts University School of Medicine. And what is 1910 Genetics? 1910 Genetics is combining artificial intelligence, computation, and biological automation to accelerate the design of small molecule and protein therapeutics. And they've built a platform for that. What are you hoping to hear from Jen today?

So I would like to give you an understanding of 1910's platform technologies, how they work, and how they address the big problems of drug discovery today. There are a lot of companies seeking to do what 1910 is doing. So I'd like to get a sense of what distinguishes them and how they turn what they're doing into a sustainable business. Well, if you're all set, let's welcome Jen to the show.

Daniel Levine (02:32.078)

So hi Jen, thanks for joining us. So before we dive into 1910 Genetics technology, I'd like to start with the company's name. So in 1910, there was something big happened in the life sciences, right? So can you tell us about that? Well, first of all, thank you for having me, Amar. I've been looking forward to this one, so thanks for having me here. 1910, I founded the company in...

2018 and I named it 1910 Genetics. And that really, at the time it was a little bit of a play on words, but the meaning of the company's name comes from the fact that, as you already know, we're an AI driven drug discovery company, but the way we select which targets to go after, which diseases to go after is really based on whether the target has a very strong

positive role in disease. And in 1910 was the year that the first patient in the United States with sickle cell disease was diagnosed. And sickle cell disease is the first disease known to scientists as having a molecular basis. So in my mind, I was thinking, oh, here's an example of the perfect disease that we should be going after because we're supremely

clear about what causes this disease and it's just a matter of bringing our AI platform to design molecules for it.

Okay, okay. So I mean, and it's been, you know, over 20, over like 114 years now, right? Since then. And especially in the last, I would say, five, six decades, there's been tremendous amount, number of drugs that have come out, right? And definitely the, what we see as drugs is very different from there, but still some of the problems with drug development persist, right? So there's still.

Daniel Levine (04:32.552)

It takes a long time for the drug development or so, for the drugs to come to the markets. So can you talk to us a bit more about what are some of the big challenges that you see in drug development right now? There's certainly several. And the biggest issue is a productivity problem, right? Productivity defined as a ratio of the outputs of new drugs that the pharma industry's bringing to market, bring to patients.

over a denominator of inputs, where inputs are a combination of human capital, financial capital, et cetera, that have gone into the process of developing those therapeutics. And what we've seen over time, over the last six, seven decades or so, is just fewer output. So the numerator is getting smaller,

or staying constant depending on the particular year you're looking at. But the denominator though is sort of ballooning. So it's costing us more. We're requiring lots and lots more money. On average, anywhere between two and two and a half billion dollars is the current cost of bringing the new medicine to market. We're taking longer to actually bring the drug to market over 10 years. And we are...

you know, it just continues to be a very failure prone process. You know, less than 10 % of novel therapeutics to start phase 1 clinical trials will actually make it to FDA approval. So it's a productivity problem, you know, and some people refer to this as Eroom's law, which is sort of the reverse of Moore's law in the semiconductor space, right? In the semiconductor space, you see the opposite phenomenon where over time, you know, that industry has seen

reducing cost of making chips relative to previous years, whereas for us in pharma, it's actually trending in the other direction. So it fundamentally is an R & D productivity problem. I think that's the issue at play here. Okay, but why is the problem there? Why is it taking that long, right, from discovering the drug's point of view? I don't know that there is one...

Daniel Levine (06:58.21)

answer, but there are a couple of different thoughts, schools of thought. I think one is we are just objectively going after much more difficult problems. You know, I trained as a classical pharmacologist at Tufts Medical and at Boston Children's Hospital of Harvard Medical School. And, you know, just looking back through just pharmacology as a discipline, you know, the very early days were, you know,

people developing things like penicillin and like much easier, well, "easier" in quotes, therapies, whereas today we're going after all kinds of modalities, right? Small molecules, large molecules, et cetera. So there's just a complexity, an increased complexity to the problems that we are tackling today. I think the other problem perhaps,

might be around just technology. We as an industry have not adopted technology to the degree that other industries have. We still very much operated in an antiquated artisanal way where much of how we go about developing, designing and developing drugs today are pretty similar to how we did them many years ago. And so I think that those two for me are probably the biggest

issues and causes for the R &D productivity problem. So in some ways, point number one can be restated as we have perhaps picked a lot of the low hanging fruit in discovery and the fruit we're trying to reach now a little higher up and we got to reach a little harder to get to them, whether it be via much more complex therapeutic modalities than say traditional small molecules or the fact that we need to be more

and adoptive of technology, are certainly far more than we are today. Okay. okay. So, you know, many companies that are built on the AI discovery platform are focused on a specific modality, whether it's small molecule or proteins or mRNA, right, or the technology. But 1910 Genetics is agnostic of drug modality, right? So can you talk a bit about why is that and how you're able to do that. that?

Daniel Levine (09:23.15)

Sure. Today we are. Today we are modality agnostic, but we didn't start that way. When we started the company five years ago, we started out building an AI platform for small molecule discovery. A couple of years into the company's existence, we started the first stages of an AI platform for large molecule, but

primarily around protein structure predictions, specifically prediction of side chain, amino acid side chain placement, once you have figured out the configuration of the backbone. And while we were sort of working in parallel, we had two separate teams, frankly, that hardly talked to each other, doing the small molecule and the large molecule work. What we started to see, and from my vantage point, when I just looked at both teams, we started to see a lot of commonalities.

And those commonalities transformed into synergies where we were better at doing certain aspects of large molecule design as a result of having done a similar thing for small molecule discovery. And so over time, what we decided to do is, as opposed to having these two parallel separate platforms, we decided to abstract, do a layer of abstraction and sort of develop

a holistic modality agnostic platform that we call input transform output today. And what we mean by that, or ITO for short, what we mean by that is regardless of the modality that we want to design, the fundamental process that we go through is the same. We start first by asking, what is the input? Which basically means what is the data that we need or we have to tackle this particular drug design challenge, right?

And we innovate in data creation by generating three different types of data. Computational data, wet lab proxy biological data, and wet lab ground truth data. Those are the three modes of data that we primarily work with to train and validate AI models. And so we start there and ask, what type of data do we need for this problem? And what is the right mix of

Daniel Levine (11:47.758)

computational and wet lab proxy and wet lab ground truth that we need. And once we answer that in the input step, then we move to the transform step. In the transform step, what we do is we look at the vast array of model architectures that we have available to us, you know, ranging from conventional machine learning all the way to deep learning, transformers, foundation models, et cetera. We are not married to any

particular model architecture, but instead we ask ourselves, what's the best architecture to use for this specific problem? And so we pick and choose, mix and match model architectures in sort of a bespoke manner to address the specific task at hand, but also to take advantage of the particular data type we have actually prepared for that problem.

So as we conclude the transform step by using the data sets, the proprietary data sets to train machine learning models, the machine learning models then go on to either design new therapeutics or optimize existing therapeutics, depending on what the actual problem statement was. And then we go on, the transform step ends with manufacturing. We actually go on to synthesize, whether its

sequences, peptide sequences, antibody sequences, or small molecule chemical matter. We synthesize them, we feed that information back in terms of synthesis, success rates, yields, et cetera, into our models, especially generative AI models, which have a tendency to sort of create and like imagine really, really wild structures that you might not actually be able to manufacture. So it's really important for us to feed that information back to the models.

And then that concludes the transform step. And then we move on to the final step, the O output. And there we are then leveraging our state of the art, robotics driven lab automation facility to design and scale up biological assays in vitro and in vivo to test those molecules, be they small or large, that we designed with the AI. So it doesn't really matter. You can come to us with a problem and you can say, hey, I have

Daniel Levine (14:04.366)

an existing monoclonal antibody, I would love to extend or to increase the half -life. Maybe it's got a one -month half -life and I want for this to have a three -month half -life. That's a problem statement. Another problem statement could be, oh, I'm targeting this kinase and it's historically been considered undruggable. I cannot find a pocket, but I think there's perhaps a cryptic pocket here somewhere. Can you design some starting Hit molecules or HIP molecules

or lead series and those are two very different problem statements. One on the Hit ID side of things and the other on the lead optimization side. One having to do with small molecules and the other having to do with large molecules. But what we're saying is we would follow the same input transfer output process in order to solve those problems. Yeah. Okay. So typically like, so I tried to maybe like take an example of either in a Rosalynd, which is your protein therapeutic

platform, or Elvis, which is for small molecule. We call them now ITO. We've replaced both of them. That idea of like separate platforms was indeed how we started. But today we have one modality agnostic input transfer output ITO platform. Gotcha. Gotcha. Okay. So in the ITO, right, if we double click on that, right, like take me through like a...

an example where like do you like usually do you start with some like you know disease signature or you you start with like some targets or you start with... like How do you like how do you usually start and how does like how does the process go and and how long does that take and and you know and how that's better than what is being done currently? So if you can take us through an example that'll be very helpful.

Right. And we are, our platform today is uniquely suited to molecular design and specifically the hit identification, hit to lead and lead optimization stages of preclinical drug discovery. So with those sort of guardrails in place, I think that what we do extremely well is first of all, we are able to

Daniel Levine (16:22.388)

design or optimize molecules far in a much quicker way than traditional approaches. We're able to leverage less humans, more machines, but really understanding that the combination of human and machine, it really is the great unlock here. We tend to, We're a pretty small team. We're less than 25 people and we're able to take on multiple targets in parallel.

And what enables us to do that is we're not as human intensive as a traditional pharma process because we definitely leverage machines and we just continue to find that nexus where human and machine work together to give you outputs that are far higher quality and outputs that you arrive at much more quickly and much more cheaply.

Um, so I gave sort of like two hypothetical examples, but they're not really a hypothetical because both were actual problem statements that um, uh, two different pharmaceutical companies brought our way. One during the COVID pandemic, they were one of the big pharma companies that were pretty, you know, far ahead in monoclonal antibody uh, development for COVID. And at the time they had um, an antibody that had, I believe it was six weeks.

half -life and they wanted to double that and and another pharma company approached us around the same time having struggled to drug a particular kinase allosterically. And for that for that latter case we were able to deliver to that pharma company in six months lead series where their own internal team hadn't been able to do that for about three to four years Okay, great great

And just for the audience, right. So the Hit is something, it's a molecule, it's a small molecule that may be binding to a protein, but then that molecule, you make changes to make it as a lead, which has a much better properties in binding the protein. And then finally you get the candidate, which is really the drug that you're going to try to do the clinical trials on. Right, right. Apologies for the jargon, but yes, a Hit is essentially that-

Daniel Levine (18:47.554)

a molecule, be it a large or small molecule, and you just want it to just do that, just hit your target. But it doesn't mean it doesn't hit other things, right? And so mice set an affinity bar of around less

than one micromolar in a biochemical assay with the free protein or less than one 10 -micromolar IC50 in a cellular assay. These are some pretty established benchmarks as to what considers

a molecule to be a binder to a particular target. But that's pretty much all you're getting at that stage. You don't know anything about that molecule sort of like promiscuity and like its tendency to have so many off target binding. You don't know anything about its drug likeness, right? How it would profile in admin assays, absorption, distribution, metabolism, excretion. You don't even know if it's toxic, right? All you know in the hit ID stage is,

here is a protein target, I've got this molecule and it binds to it. And it satisfies this binding affinity. And so that pretty much concludes your hit identification process, right? And you validate that binding and a combination of biochemical and cellular assay, and then you give yourself a thumbs up, only to then go on to the hit to lead stage and start to uncover liabilities in that molecule that might actually prevent it from ever becoming a drug.

Yeah, yeah. And so tell us about what are the distinguishing factors or the distinguishing technologies or the algorithms or the approach that 1910 is taking, right, which is making, is helping speed up this drug discovery process. Right, I think there are certainly different points of differentiation, but I'll highlight three. First is how we...

overcome the data scarcity problem that prevents the application of state of the art AI in this biotech space. Second is, you know, our robotics driven laboratory automation and how we put, you know, the AI and the data and lab automation in the same closed loop. Third is what we already spoke about, you know, the modality agnostic nature of our platform, which

Daniel Levine (21:11.502)

really truly is a differentiator. So those are like highlights, those three. And just to go into a little bit more detail, AI is fundamentally garbage in, garbage out on the data. I mean, you know this. And while our technology behemoths, like Microsoft and so on, are doing excellent work building and companies like OpenAI, obviously,

doing a lot of great work building state of the art AI models, model architectures. Many of them are not readily usable in the life sciences and in biotech specifically, in part because we don't have the scale, the breadth and depth of continuous longitudinal data

that we need to make these models work. These models tend to be very data hungry. And so one point of differentiation for 1910 is our data strategy, right? The way we overcome historically scarce wet lab ground truth data by creating two additional data types. I talked about how we use, you know,

computational simulations and generate billions of data points computationally, which we have been careful to ensure that the computational simulations are approximating biological phenomena to the best of our understanding of scientists. And so that that data, albeit synthetic, actually holds tremendous value, especially when we combine it with a second type of wet lab data that we call wet lab proxy biological data.

What we mean here is these are surrogate biological assays that we developed in the wet lab. Most of them using next generation sequencing as a way to sort of scale up your output. And what these assays are doing is they are not your ground truth validation assay. They are exactly what they sound like, a proxy assay or surrogate assays that get you about

Daniel Levine (23:21.566)

70 % of the way there in terms of the same information you'd get from a similarly low throughput ground truth assay. But with the proxy biological assay, we're able to get billions of data points in the lab. And so we combined the computational data with the wet lab proxy biological data and then used that to augment the historically scarce ground truth biological data and the.

the melding of these three multimodal data sets is what allows us to train state of the art AI models that would otherwise not work if you were only relying on the wet lab ground truth biological data as a single data source, which is what the vast majority of people in pharma are doing because they have not sort of...

understood the value of the other two data streams that I alluded to and perhaps don't have the technology that we have to actually make those other two data sets valuable in AI -driven discovery. And the second piece is around how we are...

you know, we've got a state of the art laboratory facility here in the Seaport, in the Boston Seaport. And for us, that's important, right? It's important for our two types of wet lab assays. It's important for our wet lab proxy biological data. It's important for our wet lab ground truth biological data. It's important to produce training, both training data for AI, but also to serve as validation when we actually now...

manufacture and task the output of AI, whether it be small or large molecules, having our own facility, being able to design and scale up a variety of in vitro biochemical and cellular assays, and being able to have that output in the loop, feeding back into AI models continuously, it really does speed up that iteration. And that's how we get to the faster, more quality output

Daniel Levine (25:23.406)

and in a modality agnostic way. Okay, so when we talk about the AI here, the, so there are different types of AI, right? So there's usually there's the analytical AI and then there you have, you have neural networks, but then you also have like the traditional methods, like the ensemble methods or so, and then you have the generative AI, right? So can you talk a little bit about like, and you have the three stages, right? It looks like in the eye stage, maybe there's,

I don't know, maybe more analytical AI. Is there a genetic AI that you're using? Can you tell us a little bit about like what are some of the new technologies and new algorithms that you're using? Yeah, our "I" stage is input. It's all at that stage, we're all focused on the data, right? We're trying to figure out how much computational data do we need? How much wet lab proxy biological data do we need? How much wet lab ground truth?

biological data do we need? And then as we are sort of, you know, wrapping our heads around the data strategy for that particular disease target, we're immediately then looking ahead to the transform step and thinking what types of machine learning architectures, machine learning models would best work depending on the combination of data modes, data modalities that we actually go on to use. So,

we have used and continue to use the entire gamut of models, right? Whether it's conventional classical machine learning to more state of the art, deep learning, transformer type models, large language models, small language models, foundational models, diffusion models, I mean, generative models, VAEs, the whole, like we do that whole thing. I mean, recently the team was running a study, swapping out transformers for member architecture.

We are pretty sophisticated when it comes to, you know, just a thorough understanding of these different architectures, but we don't fall in love with any one model type, right? For us, it's not, we really absolutely have to use a graph neural network because it's hot today. It really is, what question are we trying to answer here, either for our own pipeline or for this pharma company? What data, most importantly, what data, the data drives the model selection, not the other way around.

Daniel Levine (27:45.422)

We don't start by picking a diffusion model for an antibody optimization problem and then work back to figure out what data we need. No, we start by saying, what do we want to optimize with this antibody? What data do we have available to us? And given the output, we're trying to get to what model architecture is best for that. And we don't always get it right. We don't always get it right. We could start out...

using, you know, in a project actually we just submitted to Nature Communications last week, which is now undergoing review, we started out wanting to improve CNS exposure of molecules that we needed for neurological indications. And we started out using, you know, graph neural networks and then graph neural networks with attention and then attentive graph neural networks with reinforcement learning. And we just kept mixing and matching and just looking at the performance.

And so for us, it truly is an experimentation that's happening in that transform step. And we are relying on the wet lab feedback to sort of help us understand how these models are performing. And this is why in silico only plays in this space just don't make a lot of sense, right? Because you are hardly ever going to get it right the first time with your choice of model selection.

But at 1910, on any given project, my AI team is running five models in parallel for the same problem, just trying to figure out where the best leverage lies with the data set that we have uniquely created for that problem. OK. So it's pretty interesting. So it's very flexible based on the machine learning that you use is based off of the problem that you're using. Based off of the problem, based off of the data

that we are able to create and leverage. Yes, so the model selection comes later. We don't start with it. Yeah. yeah. Now in terms of the draggability, right? So there are protein families like G protein couple receptors or kinases, proteases, which are quite druggable With your approach, are you focusing on...

Daniel Levine (29:55.374)

solving some of their issues or are there also some undruggable protein families that you're looking at? Yeah, I think for us, you know, when we started out, we've worked on proteases, whether they be cysteine proteases, serine proteases, we've done quite a bit of work in the protease family, which have the, you know, the typical druggability ability problem of lack of selectivity, right? Every time you're drugging a protease, particularly cysteine proteases,

you really, really have a big off -target issue to deal with every single time. So proteases, we've done a lot of work in that target class, kinases as well. We've done work on receptors, transmembrane receptors, other types of enzymes that are not necessarily kinases or proteases. And we just continue to just expand from there. But yeah, we've transcription factors

or just sort of like scaffolding proteins. Like we've really, we've tried our hands at a variety of targets. Now, not every target on which we initiate a design campaign, do we actually then convert into a pipeline asset. For a number of reasons, we might just decide, you know, for this particular target, we just want to quickly validate a new state of the art AI architecture.

So we're just going to run one design campaign and just try to get a hit and just validate that this model architecture would work against this type of target. And then that's it. And then we actually don't proceed to then try to develop that hit into a lead and then optimize into a candidate and so on. So for us as a technology driven biotech company, we are constantly balancing drug pipeline development with platform development.

So, And sometimes we take on certain targets just purely from a platform development perspective where we want to push the platform in a new way, whether it's a new type of pocket, a new type of simulation we need to do, a new type of architecture we need to sort of validate. But we don't really have any long -term plans to advance that into a therapeutic pipeline asset.

Daniel Levine (32:16.727)

But then the flip side is also true. There are times when we're laser focused on pipeline advancement and we really are just focused on what sort of iterations of our technology really help us go from a hit to a lead and then a lead to a candidate and so on. So yeah, just really trying to balance platform development with therapeutic

pipeline development. Sometimes those two things intersect really well on the same target, on the same program. Sometimes they don't intersect well and we really just try to, you know, just move accordingly. Sure, yeah. So in terms of the business side of this, what is your business model? Is that to partner with other companies but then also develop your own pipeline?

Yeah, very much a hybrid business model where on the one hand, we first and foremost, we're a platform company and our goal is to build the definitive single most powerful AI drug discovery platform and make that available to everybody. And when I mean everybody, I mean from pharmaceutical companies, biotech companies, technology companies, government institutions, academics, even...

patients. We want to bring our technology to the world, really, truly at scale. And that's our goal. And that's what we mean when we say we want to build an infrastructure layer for tech -enabled drug discovery. And so that's goal number one. And I think goal number two is to then really hone in and double down on that infrastructure layer's ability, our infrastructure layer's ability to bring drugs to the clinic and bring clinical candidates to the market, right? And we assess that both by our own internal pipeline and how we advance in molecules from hit ID to development candidate to the clinic, et cetera. But also we assess that by the number, just a sheer number of drug development enterprises and biotech and pharma companies that our platform contributes to

Daniel Levine (34:37.853)

helping them advance molecules through the clinic and into market. So yeah, those would be the two things to highlight. Great. Now in terms of the different therapeutic areas, do you find that it's easier to develop drugs for certain therapeutic areas versus the others? So I am going to anchor on the word develop because I think of develop...

differently from design and I think a lot differently from discover. I think that at the design stage, which is just here is a target, find me a molecule that sort of binds to it or modulates it in some way. That to me is fundamentally design. I think at that stage, our platform is therapeutic area agnostic. It doesn't know. It doesn't know where you want to take that eventual molecule, right? And so it...

it means nothing to it, the actual disease. But when you arrive at a candidate series, now you have a variety of indications you could pursue. You absolutely need to have a terrific head of clinical development who is putting on their hat and saying, given, you know, psoriasis versus dementia versus fibrosis, given our belief that this

biology is implicated in these three indications, which indication should we pursue first based on XYZ parameters? So yes, from a development perspective, from a clinical development perspective, I think that really is where therapeutic area really, really matters. Because you start to get into things like, you know, orphan disease designation that you could get in some rare

therapeutic areas that you might not get in others, you start to get into some indications where the FDA is more open to leveraging telemedicine, particularly in a lot of neuro conditions, because you can have patients do a lot of patient -reported outcomes versus in others. You start to run into things like, frankly, how high a bar is the FDA going to set, right? In something like cancer,

Daniel Levine (36:58.775)

the FDA has come to accept, frankly, a much higher toxicity profile of a molecule than they would for something like lifespan extension, right? Because they're just gonna feel like, well, that's not really a big issue, right? So they're gonna set a higher bar. So you're gonna run into things like...

how big does a trial need to be, right? In the metabolic cardiovascular space, you got like 1 ,000, 2 ,000, 10 ,000 person cardiovascular studies versus in certain indications, you might just be able to get away with like a 30 person conformational trial. So disease and indication selection is absolutely critical at the development stage, but at the design stage where you're using AI to design a lot of the molecules, we're not, the platform is agnostic

to where you actually end up wanting to develop that molecule in. Sure, great, great. Now, in terms of, like you mentioned OpenAI earlier and Sam Altman, the CEO of OpenAI was one of your funders, right? like in the seed round. So how, like, can you tell us a bit about like as you're working with these investors, in addition to just the investment, how...

are you leveraging their expertise beyond just the funding? Yeah, Sam is a great investor, more importantly, a great friend. He led our seed round and he invested in our series A and we're currently raising our series B and he'll also be investing in that. He just provides just a sounding and board, you know, and on all things startup

formation, scaling, go to market, partnering strategy, et cetera. So whether you have an AI question or a question about hiring, he's just been a voice of reason for me all these years. So really appreciate him and his support and partnership. Beyond Sam, we have Microsoft's Venture Fund, which led M12, which led our...

Daniel Levine (39:18.389)

Series A back in 2021 and has positioned us to strike some important go -to -market partnerships, which we'll be announcing soon. Actually, we've got a big one coming up in January 31st. So look out for that. I think the best investors are the investors who are worth their weight in gold, just outside of just the actual

financial capital, so whether it's helping you hiring, whether it's helping you with go -to -market strategy, establishing some of those channel partnerships, thinking about ways to supercharge your technology platform above and beyond what you could do on your own as a startup with limited resources. There is not a shortage of ways that that investors could be value adds to startups and we've been fortunate to have some of those investors. that's great. That's great.

So Jen Nwanko, founder and CEO of 1910 Genetics. Jen, thank you very much for your time today. Pleasure to be here. Thank you, Amar. I enjoyed spending time with you. Thank you. Well, Amar, what did you think? Yeah, it was pretty interesting how she described the platform and then the three aspects of that and some of the interesting things she mentioned about how they are leveraging data as well as

as well as the different types of machine learning techniques to, you know, to aid their platform. Jen mentioned that 1910 didn't start out as modality agnostic, but is so today. It was interesting to hear her explain how that came about, but it strikes me as somewhat unusual in the world of AI drug development. I think of most platforms being focused on a modality. Does that make sense?

Yeah, so see, in the drug discovery process, you usually start with a target and then you're putting the data together around that. And then you're designing either a small molecule or a large molecule like a protein or antibody to target that, right. So there are there are similarities there in terms of

Daniel Levine (41:40.973)

the "I" part that she talked about the first part, like getting all the data and putting that information together. So that's definitely very similar, whether your therapeutic is a biologic or small molecule. Also the transformation part that she talked about, there's a lot of automation that's there. And again, the automation, there are parts of the automation that will be, I believe, common

in both modalities. Of course, the last part would be different, which is more with either coming up with the small molecule or the protein or antibody. That part would be different. But then in the initial aspect, I see a lot of similarities there. And then the way she described the approach, you can easily see that. One of the things she said that really stood out to me is she said the combination of

human and machine is the great unlock. Not as human intensive as traditional process, but it's human and machine working together to get higher quality drugs quicker and more cheaply. Is this something most people outside the process don't get about AI? I mean, I think there's probably a sense that we're just going to pop a quarter in the machine and it's going to spit out a drug. What don't people get about the use of AI and drug development today?

discovery? Yes, see drug discovery is a very hard problem. Essentially, there is some target protein that we are trying to target using a small molecule or a protein, a large protein. The problem is how do you design that, right? And how do you use the structure of that? And how do you how do you think about, let's say from the chemistry point of view, what are the kind of

molecules or the type of atoms that should go where. Are we really understanding the structure of the protein because the protein structure is not one, I mean it's dynamic it changes. So, I don't think the science is in a place where we know everything about a specific target. There is tremendous amount that that is for us to still learn. see I mean if you go ahead gone back about 30 years this was all done by humans but now

Daniel Levine (44:02.413)

over the last generation, a lot of the new technologies have come and then we're getting the machines involved more and more. And as time goes on, the machines are getting better. I would say that machine learning algorithms are getting better. We also are getting better data so they can do better outputs. But I believe we're still a bit far away from just the machines designing the right drug, right? And then also you have to keep in mind that a drug,

the way or like a protein, the way it is in a test tube might be different than how it's going to be in the live organism, right? So there are changes or if you just create that, I mean, there's so many differences and then the way it behaves is so different. So there are so many of these factors. And so biology is very complex, chemistry is very complex. So I think we're getting into the direction of the machines doing more and more, but we still have a long way to go.

She also touched on the issue of data scarcity and the need for data of adequate depth and scale for AI to do what it can. 1910 has a data strategy she described that's supplementing wet lab data with computational data and assay data, which she referred to as proxy data. This is generating billions of points. Does this solve the data challenge for using AI and drug discovery?

Daniel Levine (45:34.381)

It's Even to identify the right target, there are 20 ,000 genes that we have. If we look at the different types of mutations, there might be even more, right? So the usually in machine learning, you need many more samples than variables. And when we're looking at our clinical data, we're not going to have 20 ,000 patients data

for a specific disease to really trying to figure that out, to get to the bottom of exactly what are the specific protein targets that we should go after. So data is always an issue. We need a lot more data in target identification, similarly for even for chemistry, because it's a 3D conformation of the molecule that needs to fit well. So for that, you need a lot of very accurate data, which

is not necessarily there and it's hard, you know, the structural predictions, there have been tremendous progress in making the predictions about the structures, but we're not there yet, right? So even though we do have a lot of data, we still need more data. And as she said, garbage in, garbage out. You have to have a lot of data from which you can actually train the machine learning algorithm in a pretty good way. So...

what she said about data strategy, I thought that was a really good idea where they are putting together a lot of lab data and they're using that to generate hypotheses, they're using that to validate hypotheses and they're supplementing the computation with data with that. So I think that's a great idea. They're also generating a lot of the data, which is also a good idea because then you have more data to work with. So yes.

There is a lot of data available, but that's still not good enough because the problem is very hard. I think it's cliche to say someone with a hammer looks at everything as a nail. But the other thing she said that really struck me was she talked about data defining the AI model that's used. What do you think of that? Yeah, that completely makes sense. I'm trained in machine learning and AI. And that's what we always had is that don't start with

Daniel Levine (47:54.701)

a method and a machine learning method and then use it to the problem. It should be the other way around based on what the problem is and the way the data is available. Then you try to find based on that what could be the right model, right? Right? Algorithmic model for that and then you fit fit it to that. So I really like that approach. That's a flexible approach, but that's also then

you're really trying to solve the problem based on what is the best way of solving it rather than going for like the neural network every time. So I like that approach. Well, I found it a compelling conversation. I'm looking forward to the next one. Great.

Daniel Levine (48:43.917)

Thanks again to our sponsor, Agilisium Labs. Life Sciences DNA is a bi -monthly podcast produced by the Levine Media Group with production support from Fullview Media. Be sure to follow us on your preferred podcast platform. Music for this podcast is provided courtesy of the Jonah Levine Collective. We'd love to hear from you. Pop us a note at Danny at levinemediagroup .com. For Life Sciences DNA and Dr. Amar Drawid

I'm Daniel Levine. Thanks for joining us.

‍

‍

‍