3016: How Creyon Bio is Using AI to Transform Drug Development

[00:00:03] [SPEAKER_00]: What if the future of medicine lies not in discovering new drugs, but actually engineering

[00:00:10] [SPEAKER_00]: them with precision and safety at the forefront?

[00:00:14] [SPEAKER_00]: Well today I'm excited to have Chris Hart, the CEO of Creyon Bio join me on the podcast.

[00:00:20] [SPEAKER_00]: A company that is leveraging AI to revolutionise drug development.

[00:00:24] [SPEAKER_00]: And this approach has led to remarkable breakthroughs, including rapid development of a treatment

[00:00:30] [SPEAKER_00]: for a baby boy with an ultra-rare gene mutation.

[00:00:34] [SPEAKER_00]: An achievement that took 13 months from concept to dosing.

[00:00:39] [SPEAKER_00]: I want to know more about the real difference that technology made for that little boy.

[00:00:44] [SPEAKER_00]: And my guest is also going to bring a unique perspective to the intersection of AI, biotechnology

[00:00:51] [SPEAKER_00]: and personalised medicine.

[00:00:53] [SPEAKER_00]: And his experience in using AI to decode complex data and optimise chemical modification for

[00:01:00] [SPEAKER_00]: OBMs is not only innovative, but also represents a significant shift in how we think about

[00:01:05] [SPEAKER_00]: drug development.

[00:01:07] [SPEAKER_00]: So we'll talk about all that, how it could potentially shorten drug development timelines,

[00:01:13] [SPEAKER_00]: increase success rates and change the way we treat rare and complex diseases.

[00:01:19] [SPEAKER_00]: So what are the implications of AI in drug design?

[00:01:23] [SPEAKER_00]: How could this technology reshape the future of medicine?

[00:01:27] [SPEAKER_00]: These are a few things we'll explore together today.

[00:01:31] [SPEAKER_00]: Delivering daily content to 140,000 of you wonderful monthly listeners across the globe

[00:01:36] [SPEAKER_00]: is no small feat.

[00:01:38] [SPEAKER_00]: I don't want to take all the credit here because it wouldn't be possible without the backing

[00:01:41] [SPEAKER_00]: of our dedicated sponsors and partners.

[00:01:43] [SPEAKER_00]: And I also want to shine a light on the fact that legacy managed file transfer tools are

[00:01:48] [SPEAKER_00]: looking dated, they often lack the security that today's remote workforce demands.

[00:01:52] [SPEAKER_00]: And companies that continue relying on that outdated tech, they're in danger of putting

[00:01:56] [SPEAKER_00]: their sensitive data at risk.

[00:01:58] [SPEAKER_00]: Attention IT professionals, are you tired of juggling these multiple servers for secure

[00:02:03] [SPEAKER_00]: file sharing?

[00:02:04] [SPEAKER_00]: Integrated shared folders and email plus a comprehensive REST API, Kiteworks simplifies

[00:02:09] [SPEAKER_00]: your workflow.

[00:02:10] [SPEAKER_00]: For administrators, you can experience unmatched functionality and integration that traditional

[00:02:16] [SPEAKER_00]: MFT servers just can't touch.

[00:02:18] [SPEAKER_00]: Step into the future of secure managed file transfer with Kiteworks by going to kiteworks.com

[00:02:24] [SPEAKER_00]: to get started.

[00:02:25] [SPEAKER_00]: That's kiteworks.com and remember, Kiteworks is also FedRAMP, moderate authorised.

[00:02:30] [SPEAKER_00]: Thank you for your patience today.

[00:02:31] [SPEAKER_00]: This is the moment you've been waiting for.

[00:02:33] [SPEAKER_00]: It's time to welcome my guest onto the show.

[00:02:36] [SPEAKER_00]: So enough from me, let's get today's guest on.

[00:02:40] [SPEAKER_00]: So a massive warm welcome to the show, Chris.

[00:02:43] [SPEAKER_00]: Can you tell everyone listening a little about who you are and what you do?

[00:02:48] [SPEAKER_01]: Thanks, Neil.

[00:02:49] [SPEAKER_01]: Very excited to be on the show.

[00:02:50] [SPEAKER_01]: Thanks for the opportunity.

[00:02:52] [SPEAKER_01]: I'm Chris Harc, co-founder and CEO of Crayon Bio.

[00:02:55] [SPEAKER_01]: Just a bit about me.

[00:02:56] [SPEAKER_01]: I'm a scientist by training.

[00:02:57] [SPEAKER_01]: My academic training is really at the interface of biology and computer science.

[00:03:02] [SPEAKER_01]: And I've been, after I finished my doctorate work, I moved on and was, I bounced around

[00:03:08] [SPEAKER_01]: between a variety of those in biotech, pharma and even spent some time doing science policy.

[00:03:14] [SPEAKER_00]: Well, it's a pleasure to have you on the podcast with me today.

[00:03:18] [SPEAKER_00]: What put you on my radar was when I first discovered that Crayon Bio uses AI to engineer

[00:03:25] [SPEAKER_00]: oligonucleotide-based medicines or OBMs.

[00:03:28] [SPEAKER_00]: There's so much hype around AI and things now, but how are you seeing AI enhancing the

[00:03:34] [SPEAKER_00]: drug development process, particularly in ensuring safety right from the outset?

[00:03:39] [SPEAKER_00]: Because as I said, we hear a lot about the hype, but this is real problems we're solving here.

[00:03:44] [SPEAKER_01]: Indeed.

[00:03:45] [SPEAKER_01]: And just to take a step back, because I think it'll help contextualize how we're using AI

[00:03:49] [SPEAKER_01]: and machine learning in drug development at Crayon, it's pretty good just to give the

[00:03:55] [SPEAKER_01]: listeners a little bit of a background as to what oligonucleotide-based medicines are.

[00:03:59] [SPEAKER_01]: Sometimes I'll refer to them as OBMs because oligonucleotide-based medicines is a bit of

[00:04:03] [SPEAKER_01]: a long phrase, but OBMs are, the way to think about them is they're effectively DNA-like

[00:04:10] [SPEAKER_01]: molecules, short DNA-like molecules that we chemically modify to improve their drug-like

[00:04:15] [SPEAKER_01]: properties.

[00:04:16] [SPEAKER_01]: And one of the things that makes them remarkable as a class of drugs is that they are foundationally

[00:04:22] [SPEAKER_01]: different than other types of molecules.

[00:04:24] [SPEAKER_01]: The first big difference is that they're polymeric, so they're different than small molecules,

[00:04:28] [SPEAKER_01]: but even amongst polymeric drugs like peptides or proteins, OBMs are special in a different

[00:04:34] [SPEAKER_01]: sense, which is both peptides or proteins or antibody drugs.

[00:04:38] [SPEAKER_01]: They all interact with things in the cell to have pharmacological effect through the

[00:04:42] [SPEAKER_01]: complex interactions of really how the drug interacts with these biomolecules.

[00:04:47] [SPEAKER_01]: And OBMs foundationally work by binding to the information molecules in the cell, either

[00:04:53] [SPEAKER_01]: RNA or DNA, and in our case, largely RNA, through Watson-Crick-Franklin hybridization,

[00:05:00] [SPEAKER_01]: which is A's to T's and G's to C's, as we've all learned in probably high school biology,

[00:05:04] [SPEAKER_01]: which makes them effectively quote-unquote information drugs.

[00:05:07] [SPEAKER_01]: So we're no longer trying to sort of make and connect really complicated puzzle pieces

[00:05:12] [SPEAKER_01]: where we have to understand the structure of protein and how molecules can interact with

[00:05:17] [SPEAKER_01]: that complex structure in the complexity of a cell environment, but rather we can leverage

[00:05:21] [SPEAKER_01]: the information system that the cell uses all the time.

[00:05:24] [SPEAKER_01]: And you put that in the context of where we live today in a sort of an arc of technology,

[00:05:30] [SPEAKER_01]: about 20 years ago, we sequenced the first human genome and it was a race between a private

[00:05:36] [SPEAKER_01]: sector and the public sector and they declared victory, which is pretty amazing.

[00:05:41] [SPEAKER_01]: And what that victory brought us from a science perspective, especially medicine at large,

[00:05:46] [SPEAKER_01]: we now can unravel the genetic and molecular basis of disease better than ever before.

[00:05:53] [SPEAKER_01]: And we know often what genes need to be controlled to have therapeutic benefit.

[00:05:58] [SPEAKER_01]: And all of those allow you to have a direct path to making that control.

[00:06:03] [SPEAKER_01]: The other piece of the world of oligonucleotide-based medicines is actually a pretty

[00:06:07] [SPEAKER_01]: mature field. It's been sort of in the shadows of drug development for a while, but the first

[00:06:14] [SPEAKER_01]: seminal experiments that showed this was a feasible idea came into the light about 40 years ago.

[00:06:20] [SPEAKER_01]: And over the last 40 years, what's really emerged is that you can make good drugs out of these

[00:06:25] [SPEAKER_01]: chemically modified DNA-like molecules. They can be blockbuster drugs, but what has been

[00:06:31] [SPEAKER_01]: slowing the adoption of this medicine is that the hope that it would be a really fast and easy

[00:06:38] [SPEAKER_01]: translation from information about an RNA molecule that you want to control to a drug,

[00:06:42] [SPEAKER_01]: because just write your first compliments, it turns out that's easy.

[00:06:46] [SPEAKER_01]: But when we put these chemical modifications onto the types of medicines we make,

[00:06:50] [SPEAKER_01]: these molecules often end up being toxic or inactive. And that complexity is built into

[00:06:57] [SPEAKER_01]: the frame of how do you assemble these polymers? And what we recognized at Crown is these are

[00:07:03] [SPEAKER_01]: polymeric drugs, which means that like other polymers, we should be able to engineer these

[00:07:07] [SPEAKER_01]: compounds like we engineer every other polymer that we depend on every day.

[00:07:11] [SPEAKER_01]: But the problem is no one had created the right data to really understand the complexity of how

[00:07:17] [SPEAKER_01]: the different components would interact. And when you do start to tease that information out by

[00:07:22] [SPEAKER_01]: creating the right data, you also recognize that this is a very complicated landscape

[00:07:27] [SPEAKER_01]: with many intricacies and complexity. And this is where AI and machine learning comes to bear and

[00:07:33] [SPEAKER_01]: has just huge potential to help really understand these complex data spaces. And that's what we've

[00:07:37] [SPEAKER_01]: leveraging AI for is really how do we first create the right data and then we can feed

[00:07:42] [SPEAKER_01]: into these models. And these models can help us understand the data so that we can walk out

[00:07:46] [SPEAKER_01]: not with simply heuristics, but rather with engineering insights so that we can create

[00:07:51] [SPEAKER_01]: these medicines much more facile than we could if we were trying to use traditional trial and

[00:07:57] [SPEAKER_00]: error-based methods. And as someone that has lost a child to a rare disease, your recent success

[00:08:04] [SPEAKER_00]: in treating a baby boy with an ultra-rare disease is nothing short of remarkable. So can you tell

[00:08:10] [SPEAKER_00]: me a little bit more about this case and also the role that AI played in developing the treatment?

[00:08:16] [SPEAKER_01]: Of course. Yeah, this was a pretty amazing story. And it really is a hallmark of where we're headed

[00:08:23] [SPEAKER_01]: in terms of modern biology and what it has to sort of offer to this notion of translating the

[00:08:29] [SPEAKER_01]: genetic understanding of disease to novel therapies. And in this case, this is work we did in

[00:08:34] [SPEAKER_01]: the TNPO2 Foundation. And the TNPO2 Foundation was established by a family whose son was born,

[00:08:43] [SPEAKER_01]: unfortunately, with a variant in the gene TNPO2, transportin-2. Just to give you the context,

[00:08:50] [SPEAKER_01]: transportin-2 is a gene that encodes a protein that's important for cellular function, in particular

[00:08:56] [SPEAKER_01]: neurons, and it works to keep things in and out of the nucleus, which keeps neurons working well.

[00:09:01] [SPEAKER_01]: But the story of our collaboration really started when Leo was born and the family and the clinicians

[00:09:10] [SPEAKER_01]: realized he wasn't thriving, and they had an opportunity to get him sequenced. And they

[00:09:14] [SPEAKER_01]: sequenced his whole genome, and they diagnosed him with a very rare disease. In fact, the disease

[00:09:21] [SPEAKER_01]: had only been identified in one other individual in the entire world ever before. And it was, in

[00:09:27] [SPEAKER_01]: fact, a variant that was a single nucleotide polymorphism, which is to say a single A, C,

[00:09:33] [SPEAKER_01]: T, or G was changed from what it is in all of us to a different base in this gene, a one in three

[00:09:40] [SPEAKER_01]: billion error, if you will. And because we live in the world of modern biology, we're able to look

[00:09:45] [SPEAKER_01]: at that data and understand that the variant there was likely causing that gene to be dysfunctional.

[00:09:53] [SPEAKER_01]: And it was likely causing a production of a protein that wasn't just broken, but rather

[00:09:58] [SPEAKER_01]: causing additional toxicity. It'd be just an analogy I sometimes like to use is if you made

[00:10:05] [SPEAKER_01]: tires at a factory, and in this case, the tires are proteins coming off of a gene,

[00:10:11] [SPEAKER_01]: and your factory made 50% of the tires to be broken, right? They just couldn't hold air.

[00:10:18] [SPEAKER_01]: And then you shipped all your tires off to the car manufacturer. Pretty much every single car

[00:10:24] [SPEAKER_01]: would come out of the lot broken because it's really unlikely if 50% of your tires are broken,

[00:10:30] [SPEAKER_01]: that you'd end up with four tires that were working. And this is effectively what was

[00:10:34] [SPEAKER_01]: happening inside of Leo cells is that where they were making 50% of this TMPO2 protein was bad,

[00:10:40] [SPEAKER_01]: and it was causing aggregations of effectively this misfunction to build up. But the sort of

[00:10:45] [SPEAKER_01]: tire analogy it holds is also the sort of analogy to tell you how you can fix this, which is if you

[00:10:51] [SPEAKER_01]: just got rid of before you sent those tires out to the car manufacturer, the bad tires. If we just

[00:10:57] [SPEAKER_01]: got rid of the bad TMPO2 gene, the good gene, as long as you had enough of it would suffice.

[00:11:02] [SPEAKER_01]: In the car analogy, maybe you'd make slightly fewer cars, but you'd still have functional cars.

[00:11:07] [SPEAKER_01]: And in this case, also a testament to modern biology, we knew that there are people in the

[00:11:12] [SPEAKER_01]: world, because we've sequenced millions of people now that actually live just fine with only one

[00:11:17] [SPEAKER_01]: functional copy of TMPO2. So that gave us confidence that there'd be something to do.

[00:11:23] [SPEAKER_01]: And it gave the family confidence as well. So they were diagnosed when Leo was just three months old.

[00:11:28] [SPEAKER_01]: And then about seven months later, they found us and they found us having this information in their

[00:11:34] [SPEAKER_01]: backpack, if you will. And when they found us, we were also in a fortunate place to be able to help.

[00:11:39] [SPEAKER_01]: We had just finished putting together the final touches on our first iteration of our platform.

[00:11:44] [SPEAKER_01]: When I say our platform, this is a sort of a collection of data, machine learning tools,

[00:11:49] [SPEAKER_01]: and assays that allow us to very rapidly go from a gene you want to modulate to a drug that would be

[00:11:56] [SPEAKER_01]: expected to be safe and active in people. And the tools we need to validate that assertion would

[00:12:02] [SPEAKER_01]: be right. So when they came to us, we said, yes, we'd help. And ultimately what we ended up doing

[00:12:08] [SPEAKER_01]: is we used the first iteration of this platform, which worked as we had all hoped it would. And

[00:12:13] [SPEAKER_01]: it allowed us to go from really whiteboard to a lead drug that we were ready to scale up for

[00:12:20] [SPEAKER_01]: carrying all the information you need to sort of get approval from the US FDA to dose the patient.

[00:12:26] [SPEAKER_01]: And that first worked from engineering to validating work in both patient fibroblasts

[00:12:32] [SPEAKER_01]: and edible models to get to a place where we're ready to scale to get to a drug product, if you

[00:12:37] [SPEAKER_01]: will, all took us five months. And for those who aren't in the sort of drug development world,

[00:12:42] [SPEAKER_01]: that's remarkably fast. And on top of that, we did it close to break us speeds, but also we ended up

[00:12:48] [SPEAKER_01]: with, we tested in total 96, but of the first 48 we tested, we had over eight leads that came out

[00:12:54] [SPEAKER_01]: that were all like great drugs. So we had to choose one. We did, we scaled it up for what we call good

[00:13:00] [SPEAKER_01]: manufacturing processes, which you need to make drug products. And then we did the GLP studies,

[00:13:05] [SPEAKER_01]: which are good laboratory practices. And then we filed for an application with a clinician that

[00:13:10] [SPEAKER_01]: was going to treat Leo and an investigator initiated IND and the FDA gave us the go ahead

[00:13:15] [SPEAKER_01]: to treat Leo. And we started treating him last July. So he's been on drug now for over a year.

[00:13:21] [SPEAKER_01]: And I'm super excited to report that in this process of engineering a drug, validating,

[00:13:27] [SPEAKER_01]: and then getting approval to dose a patient all at breakneck speeds of less than 13 months to go

[00:13:31] [SPEAKER_01]: from whiteboard to dosing. We also ended up making a drug that's having significant benefit

[00:13:37] [SPEAKER_01]: for Leo. He had, even in the timeframe in which we met the family, his development of milestones

[00:13:43] [SPEAKER_01]: started to slip and his severity of seizures started to get worse. So by the time we dosed him,

[00:13:49] [SPEAKER_01]: he was having seizures, often seizures that would stop him from breathing half a dozen to a dozen

[00:13:54] [SPEAKER_01]: times a week, if not more frequently. And he had lost his ability to really engage with the world.

[00:14:00] [SPEAKER_01]: He just couldn't control himself. He couldn't focus his eyes and really was sort of non-responsive

[00:14:06] [SPEAKER_01]: to the environment he was in. And then we started dosing him. And then we started at a sub-therapeutic

[00:14:12] [SPEAKER_01]: dose to make sure that we're being cautious in case there were some adversity that we did not

[00:14:17] [SPEAKER_01]: pick up in our preclinical work. Fortunately, that did not happen. And by the time we got to the fall,

[00:14:22] [SPEAKER_01]: we were able to get to a therapeutic dose and we started to see real benefit. There was a precipitous

[00:14:26] [SPEAKER_01]: drop in his seizure frequencies. We saw milestones regain and he's clearly better off than he was

[00:14:34] [SPEAKER_01]: not on drug. And that is continuing to be an improvement that we have seen throughout the

[00:14:40] [SPEAKER_01]: year. And that actually just gave us the sort of go ahead to increase the dosing frequency,

[00:14:46] [SPEAKER_01]: even because the drug has been so well received. The sort of the high notes there is that on top of

[00:14:52] [SPEAKER_01]: making a drug that's having tremendous benefits for this patient, we also demonstrated the platform

[00:14:59] [SPEAKER_01]: allows us to make safe and effective drugs hyper-efficiently. And it allowed us to make

[00:15:04] [SPEAKER_01]: really the first lox nucleic acid antisense oligo that's a real selective drug that's ever been

[00:15:10] [SPEAKER_01]: administered to a patient. And that's actually a pretty big deal because LNAs, these are chemical

[00:15:15] [SPEAKER_01]: modifications that were discovered in the 2000s-ish. And people hadn't been able to apply

[00:15:20] [SPEAKER_01]: them even though they improved the activity of drugs a significant tenfold. People have been

[00:15:25] [SPEAKER_01]: able to apply them in the clinic as readily as they would have liked to. In fact, there was a

[00:15:29] [SPEAKER_01]: splash of publications that came out in the 2010s era associated with this type of chemistry where

[00:15:35] [SPEAKER_01]: people said they're just flat out toxic, which was 100% wrong. These drugs are not toxic,

[00:15:41] [SPEAKER_01]: car blank. They're only toxic if they're put on the wrong sequences and in the wrong configuration.

[00:15:46] [SPEAKER_01]: What the platform we developed is it allows us to actually leverage these sorts of modern

[00:15:51] [SPEAKER_01]: chemistries in a way where we can control the safety from an engineering principle point of

[00:15:56] [SPEAKER_01]: view and not rely on trial and error screening, which allows you where you fail to actually be

[00:16:00] [SPEAKER_01]: able to identify these safe compounds. So this drug is actually a first of many kinds. And it's

[00:16:06] [SPEAKER_01]: also the first of many that we'll be able to create off of the platform we developed.

[00:16:10] [SPEAKER_00]: Absolutely phenomenal what you've achieved here and a big shout out to Leo and best wishes from

[00:16:16] [SPEAKER_00]: me to him and his family. And I'm curious, how does CryonBio's approach to drug development,

[00:16:23] [SPEAKER_00]: how does it differ from, let's say more traditional methods? And what significant

[00:16:27] [SPEAKER_00]: advantages does AI bring to identifying safe and effective medicine?

[00:16:33] [SPEAKER_01]: That's great. So I think that the sort of the way you can think about is that we're truly

[00:16:37] [SPEAKER_01]: engineering rather than discovering molecules. Traditional drug development has always been

[00:16:43] [SPEAKER_01]: predicated on this idea, which is it's hard to find bioactive molecules. And that sort of makes

[00:16:52] [SPEAKER_01]: sense here. The cells are robust, finding a molecule that are interact very precisely with

[00:16:57] [SPEAKER_01]: some biological process to have therapeutic benefit. That's challenging. We've gotten better

[00:17:02] [SPEAKER_01]: over the years with high throughput screening and now generative AI methods to sort of identify

[00:17:07] [SPEAKER_01]: these complex interactions that may be helpful in driving these things. But at the end of the day,

[00:17:11] [SPEAKER_01]: that entire workflow is built on this notion that the serendipitous moment for drug discovery

[00:17:17] [SPEAKER_01]: is when you can find a molecule that interacts with another molecule to have benefit. And that

[00:17:23] [SPEAKER_01]: interaction is the thing you drive for. And in other words of saying that they look for activity

[00:17:27] [SPEAKER_01]: first. And then once you show something's active, you'll go forward and you'll say,

[00:17:32] [SPEAKER_01]: great, I hope this thing is safe. And often it's not. And then often you'll spend the next

[00:17:39] [SPEAKER_01]: iterations, which actually turns out to be sometimes years to decades and sometimes it'll

[00:17:44] [SPEAKER_01]: cost hundreds of millions of dollars to iterate on that process hoping you can find something that

[00:17:49] [SPEAKER_01]: would be well tolerated enough to go with the clinic and retain the same activity.

[00:17:53] [SPEAKER_01]: What we've built is something that allows us to use oligos, that allow us to sort of avoid

[00:17:57] [SPEAKER_01]: and skirt the problem of having to figure out how we're going to interact with complex

[00:18:01] [SPEAKER_01]: three-dimensional structures. We just interact in the way that information molecules in the cells

[00:18:06] [SPEAKER_01]: exist through hybridization. So, that's the first thing. But the second thing is how do we make

[00:18:11] [SPEAKER_01]: sure... So, activity was never really a major problem or binding was never really a big

[00:18:16] [SPEAKER_01]: problem in the space, but the sort of assurance of safety was very much the challenge in the space.

[00:18:21] [SPEAKER_01]: And what we've built out is a way of actually understanding from data and then the machine

[00:18:26] [SPEAKER_01]: learning models that allow us to then deploy them with sort of engineering crispness where

[00:18:30] [SPEAKER_01]: and how we should put chemical modifications in these molecules based on the sequence of them

[00:18:35] [SPEAKER_01]: such that we can have a very high expectation that they would be well tolerated.

[00:18:40] [SPEAKER_01]: That has huge implications right now. We've cut out this chunk of early drug development,

[00:18:45] [SPEAKER_01]: which is both failure prone, costly, and slow and unpredictable largely. So,

[00:18:50] [SPEAKER_01]: when we think about the process, one of the things that we're systematically trying to do at Cryon

[00:18:54] [SPEAKER_01]: is erode the cost and time it takes to create new medicines. And this was the first thing we had to

[00:19:01] [SPEAKER_01]: solve to make sure we could expedite the process of going from target ID to a lead that we could

[00:19:07] [SPEAKER_01]: push into. And AI was a critical piece of this because the complexity of the design world

[00:19:13] [SPEAKER_01]: associated with oligos is just... It's immense, and the interactions of the components that make

[00:19:18] [SPEAKER_01]: up oligos is also highly complicated. So, this would be an untenable task if you were sort of

[00:19:26] [SPEAKER_01]: trying to write down simple heuristics that you could understand and sort of just draw on like

[00:19:31] [SPEAKER_01]: a drug should only have three of these in a row or two of that. That's inadequate. The way you

[00:19:37] [SPEAKER_01]: have to think about this is in a much more holistic way, which allows you to take into

[00:19:42] [SPEAKER_01]: the complex correlations that exist between every aspect of the molecule.

[00:19:45] [SPEAKER_00]: And I think the intersection of AI and medicine is a rapidly evolving field. And

[00:19:50] [SPEAKER_00]: I'm curious from what you're seeing, what are the particular trends or maybe even advancements in

[00:19:56] [SPEAKER_00]: this area that excite you the most? And how do you see them shaping the future of biotech?

[00:20:02] [SPEAKER_00]: I understand it is very early, but what particularly excites you at the moment?

[00:20:06] [SPEAKER_01]: Yeah, it's a great question. I think AI is... I'm excited about many aspects of

[00:20:10] [SPEAKER_01]: the clinical clearing. But in biotech, one of the challenges that you face is that

[00:20:16] [SPEAKER_01]: data is often sparse. So, in the near term, a lot of the things that you think about where

[00:20:21] [SPEAKER_01]: we're going to see advances are going to be in places where we have access to data and where

[00:20:26] [SPEAKER_01]: there's opportunities to leverage that data in novel ways to gain insights. That'd be difficult

[00:20:31] [SPEAKER_01]: for us to do as mere mortal brains because we can't hold that corpus of information in our heads

[00:20:37] [SPEAKER_01]: as well. So, when I think about that, I think about what and where are you really excited

[00:20:42] [SPEAKER_01]: because it's going to have huge impacts in places like understanding disease at a resolution where

[00:20:48] [SPEAKER_01]: we could actually have preventative bell towers, if you will, to let us know that we're walking

[00:20:53] [SPEAKER_01]: down places without having expensive interventions with biomarkers. And a lot of people think about

[00:20:58] [SPEAKER_01]: biomarkers from what sort of things you can measure out of the blood that would be a signal

[00:21:01] [SPEAKER_01]: of disease. I think we see, I'm wearing a smart watch, Neil, I can't see you, but you may be wearing

[00:21:07] [SPEAKER_01]: a smart watch. The amount of information that we can now get off of these digital tools that allow

[00:21:12] [SPEAKER_01]: us to really understand at both a temporal basis, at an individualized basis, and at a cadence

[00:21:20] [SPEAKER_01]: that allows us to see perturbations from normal and then connect to what would be expected for

[00:21:25] [SPEAKER_01]: you personally because we're all somewhat different based on the corpus of data that

[00:21:29] [SPEAKER_01]: people can collect. I'm very excited that we'll be able to sort of have these biomarkers that come

[00:21:33] [SPEAKER_01]: out, especially digital biomarkers that allow us both to have earlier diagnosis that things are going

[00:21:38] [SPEAKER_01]: amiss so we maybe can have earlier interventions. But in addition, that clarity also allows us when

[00:21:44] [SPEAKER_01]: we're thinking about developing novel therapies to understand in more clarity sort of when an

[00:21:50] [SPEAKER_01]: intervention is having benefit for patients in a way that will allow us to have hopefully perhaps

[00:21:56] [SPEAKER_01]: trials that, especially in the rare communities, don't have to rely so deeply on placebo controls

[00:22:02] [SPEAKER_01]: or we can get away with shorter time scales or we can get away with even just higher resolution

[00:22:07] [SPEAKER_01]: of which patients under what conditions would have benefit to the drugs we're pushing forward.

[00:22:12] [SPEAKER_01]: Along those lines, that falls into allowing us to do things like build out digital twins or having

[00:22:17] [SPEAKER_01]: better diagnoses and sort of I think that'll also fall forward into understanding really the systems

[00:22:24] [SPEAKER_01]: medicine arena around disease pathologies. I completely agree with you and I'm currently testing a smart ring

[00:22:31] [SPEAKER_00]: very similar to the smart watch there. It really got me thinking about how we're moving towards more

[00:22:37] [SPEAKER_00]: proactive approach to healthcare rather than the reactive approach of waiting for something to go

[00:22:42] [SPEAKER_00]: wrong. For you guys, I mean with the ability to engineer drugs for safety first, how does AI

[00:22:49] [SPEAKER_00]: influence the speed and efficiency of bringing these new medicines to market? I know we've talked

[00:22:54] [SPEAKER_00]: about this a little already but is there anything else you can share around that on influencing that

[00:22:59] [SPEAKER_01]: speed and efficiency? Yeah, we often think about it just in terms of this is not new in the field

[00:23:04] [SPEAKER_01]: obviously but thinking that one of the challenges in drug development generally is that

[00:23:10] [SPEAKER_01]: from idea to drug in the market time frames is both. One way of thinking about it is that that's

[00:23:16] [SPEAKER_01]: the biggest problem which is it costs too much to make too few drugs. What's the driver of that?

[00:23:20] [SPEAKER_01]: It's really we have a lousy probability of success from project initiation to making a drug. We think

[00:23:26] [SPEAKER_01]: a lot of the work that we're doing is really sort of foundationally pushing them forward on the way

[00:23:31] [SPEAKER_01]: of how do we improve the probability of success in that process and the way that we sort of started

[00:23:36] [SPEAKER_01]: where we are and the reason that we started where we were was that when we thought about it, what

[00:23:43] [SPEAKER_01]: would be the biggest impact on improving probability of success and hopefully sort of mitigating costs

[00:23:48] [SPEAKER_01]: in terms of time and real costs to make drugs so that we can get rid of this problem. If you look

[00:23:53] [SPEAKER_01]: across the industry and ask why do drugs fail, there's really in broad strokes there's a handful

[00:23:58] [SPEAKER_01]: of reasons not many. One is they fail because they have safety liabilities. They fail because

[00:24:03] [SPEAKER_01]: that the therapeutic hypothesis is wrong. You're modulating the wrong thing. It just doesn't help.

[00:24:08] [SPEAKER_01]: You made the wrong drug. It doesn't do what you thought it was going to do and the last one which

[00:24:13] [SPEAKER_01]: people don't always think about is that your drug doesn't actually get to where it needs to go

[00:24:17] [SPEAKER_01]: to have a therapeutic benefit. So when we think about what we've built is we've built a platform

[00:24:22] [SPEAKER_01]: that really foundationally resolves the safety liabilities. How do we connect data together from

[00:24:27] [SPEAKER_01]: the corpus information we create so that we truly have for every drug we make an improved probability

[00:24:33] [SPEAKER_01]: of success in terms of safety? I think we're on track doing that. We're creating new data every

[00:24:37] [SPEAKER_01]: day and our platform allows us to integrate that data in ways that others just can't.

[00:24:41] [SPEAKER_01]: And it's truly a platform. It's not a set of learned heuristics and it's not hope that

[00:24:46] [SPEAKER_01]: our magic chemistry is going to make our drugs better but rather it's data-driven

[00:24:51] [SPEAKER_01]: systems that allow us to truly understand which chemistries under which contexts are going to

[00:24:55] [SPEAKER_01]: have benefit to improving safety. The second thing I mentioned was we're going after the wrong

[00:25:00] [SPEAKER_01]: therapeutic hypothesis. It's been long established in the field that if you're targeting the genetic

[00:25:04] [SPEAKER_01]: basis of disease, you actually have a better chance of success generally because you're

[00:25:09] [SPEAKER_01]: actually going after something that has sort of a causal error if you will. So oligos give us that

[00:25:14] [SPEAKER_01]: advantage already and we know because oligos all work through generally the same molecular actions

[00:25:19] [SPEAKER_01]: we're unlikely to create what are the wrong drug. It'll certainly do what we want it to do at a

[00:25:25] [SPEAKER_01]: molecular level. So those are the first three things that I think we've sort of been addressing

[00:25:28] [SPEAKER_01]: with the platform and our choice of substrate if you will. And the last one delivery problems I

[00:25:32] [SPEAKER_01]: didn't talk about but this is really the sort of the other Achilles heel of the industry is that

[00:25:36] [SPEAKER_01]: oligos don't have broad tissue access by default. So what we've also built forward is a way for us

[00:25:42] [SPEAKER_01]: to improve the distribution of oligos into tissues they'd otherwise not be able to have access to.

[00:25:49] [SPEAKER_01]: And we do that by leaning on our core competency around engineering nucleic acids and we use a

[00:25:55] [SPEAKER_01]: set of technologies called aptamers. Aptamers are effectively chemically modified nucleic acids

[00:26:00] [SPEAKER_01]: that rather than interacting with the sort of their targets through Watson-Crick-Franklin

[00:26:05] [SPEAKER_01]: hybridization information code if you will, they actually do form complex three-dimensional

[00:26:10] [SPEAKER_01]: structures that can bind to receptors on cells. And what I think is interesting about that is

[00:26:15] [SPEAKER_01]: that we can use them but unlike trying to do that to actually control the molecule,

[00:26:20] [SPEAKER_01]: we're only looking to bind to that molecule and we're only looking to bind to that molecule

[00:26:25] [SPEAKER_01]: long enough to actually bring our payload that can control the gene expression into a cell which

[00:26:31] [SPEAKER_01]: is actually a much simpler problem than if you're trying to do that sort of interaction to create a

[00:26:36] [SPEAKER_01]: novel therapy. And what we've built is a set of tools that allow us to do that very efficiently.

[00:26:41] [SPEAKER_01]: So now we've solved what I think are the foundational pillars that often drive

[00:26:46] [SPEAKER_01]: drug failures improving and by improving the probability of success, our hope is that we'll

[00:26:50] [SPEAKER_01]: have an outstretched sort of capacity to improve the ways in which we can create novel medicines

[00:26:57] [SPEAKER_01]: for both rare diseases, ultra-rare diseases like we did for Leo but even common diseases as we push

[00:27:03] [SPEAKER_01]: forward as the processes are the same regardless of whether you're treating a patient for

[00:27:07] [SPEAKER_01]: as a bespoke variant that's driving their disease or you're treating a disease that's

[00:27:13] [SPEAKER_00]: fairly common in the population. And I'm conscious we might be making it sound very easy but I'm sure

[00:27:18] [SPEAKER_00]: there's been a lot of challenges along the way. So what initial challenges did you face when

[00:27:24] [SPEAKER_00]: integrating AI with traditional biotechnological practices and how have you at CrayonBio

[00:27:31] [SPEAKER_00]: overcome some of these obstacles? Any challenges you can share there?

[00:27:35] [SPEAKER_01]: I think that I alluded this before, the biggest challenge I think it is twofold. One is data

[00:27:40] [SPEAKER_01]: and the second everything we do is dependent upon having an exceptional team of people

[00:27:44] [SPEAKER_01]: and the work that we do is challenging because it actually spans a broad set of domain expertise.

[00:27:52] [SPEAKER_01]: So finding people that are domain experts and then making sure that we are able to assemble

[00:27:58] [SPEAKER_01]: teams that can work together well and span these disparate domains that need to be spanned

[00:28:03] [SPEAKER_01]: is something that we've spent a fair amount of time cramming. I think the team's spectacular

[00:28:06] [SPEAKER_01]: and we're constantly working in frames where we have deeply quantitative scientists working with

[00:28:11] [SPEAKER_01]: bench scientists and these dry scientists and wet scientists working together with

[00:28:16] [SPEAKER_01]: across domains whether it's chemistry, pharmacology, computer science, mathematics,

[00:28:20] [SPEAKER_01]: physics. That teaming is critical. The second challenge that I think is obvious is that data

[00:28:26] [SPEAKER_01]: is king here. So we spent the last five, almost five years now at Crayon, the vast majority of

[00:28:32] [SPEAKER_01]: that time was thinking about how do we create the right data to build models because what we

[00:28:36] [SPEAKER_01]: recognize is that retrospective data for many AI applications like if you want to build a large

[00:28:42] [SPEAKER_01]: language model to predict your next word or your next sentence as you're writing or help with a

[00:28:47] [SPEAKER_01]: variety of things to render pictures, that's great because you have terabytes and terabytes of data.

[00:28:53] [SPEAKER_01]: If you want to engineer what should be your next drug, you don't have that luxury because there's

[00:28:59] [SPEAKER_01]: just not that much interconnected data. So we had to think very deeply about how do you think about

[00:29:04] [SPEAKER_00]: creating smart data, not just big data. I think the other area we've not discussed is how regulated

[00:29:10] [SPEAKER_00]: the industry is. It's an incredibly highly regulated industry. So how do you ensure that

[00:29:16] [SPEAKER_00]: the use of AI in drug development aligns with some of those regulatory requirements, particularly in

[00:29:22] [SPEAKER_00]: the context of developing treatments for rare and ultra-rare diseases? Any obstacles here or

[00:29:29] [SPEAKER_00]: has the community been largely embracing of it? I'm pretty excited about the tails of the sail,

[00:29:35] [SPEAKER_01]: if you will, from what regulators are working on. It's an emerging space in terms of how does AI

[00:29:40] [SPEAKER_01]: feed into the drug development, but both FDA and agencies at the UK and Europe are actively

[00:29:46] [SPEAKER_01]: pursuing both seeking public comments and feedback from industry about how can we leverage AI

[00:29:53] [SPEAKER_01]: optimally to help sharpen our regulatory processes. So I think we're on a path here,

[00:29:58] [SPEAKER_01]: which is good. It's going to be a true collaboration between regulators and clinicians

[00:30:03] [SPEAKER_01]: and patients in the industry to sort of figure out what is the best way so that we can do what

[00:30:08] [SPEAKER_01]: we need to do, which is assure safety and efficacy of these types of medicines we make while also

[00:30:13] [SPEAKER_01]: leveraging the tooling that's coming up from novel science optimally. So I think that's an evolving

[00:30:18] [SPEAKER_01]: space that's moving in and hopefully what will be a continually puggy place. And then I think from

[00:30:24] [SPEAKER_01]: the sort of regulatory space around rare versus ultra-rare versus common diseases, that's also a

[00:30:30] [SPEAKER_01]: place that I think that regulators have been very forward thinking in many spaces where they've made,

[00:30:36] [SPEAKER_01]: I would say, simplifying processes to enable rare disease development where it'd be otherwise

[00:30:42] [SPEAKER_01]: untenable. The work we did with the TMPO2 Foundation treating Leo, we were able to leverage,

[00:30:47] [SPEAKER_01]: for instance, the NF1 guidance that the FDA put forward to allow us to sort of have an

[00:30:53] [SPEAKER_01]: expedited path through the agency through an investigator-initiated application. I think

[00:30:58] [SPEAKER_01]: there's work to be done there to broaden that out to sort of see how we can connect those advances to

[00:31:05] [SPEAKER_01]: industry such that there's viable commercial models to allow for stronger innovation. But I think that

[00:31:13] [SPEAKER_01]: momentum in that space is going in the right direction. And we're always excited to work

[00:31:18] [SPEAKER_00]: with policymakers and regulators as we do the work we do. And looking ahead, what are Cray and Bio's

[00:31:25] [SPEAKER_00]: goals for the next few years? How do you see AI continuing to transform the landscape of drug

[00:31:31] [SPEAKER_00]: development, personalized medicine, and so much more? There's so many big opportunities there,

[00:31:35] [SPEAKER_00]: but how do you see that taking shape? What are your big goals there? Our sort of next couple

[00:31:40] [SPEAKER_01]: of years will hopefully be marked as successes of taking the platform we've developed and using it

[00:31:45] [SPEAKER_01]: to both advance our internal programs, and we also have partnerships with other biotech and

[00:31:51] [SPEAKER_01]: pharma companies that we're excited to work through to sort of advance compounds into the

[00:31:55] [SPEAKER_01]: clinic, not just for NF1 diseases but also for common diseases. As those go forward,

[00:32:01] [SPEAKER_01]: we're very excited to have an increasing set of data that connects the work that we've been able

[00:32:05] [SPEAKER_01]: to do to really engineer compounds for safety as far as we can tell with preclinical readouts.

[00:32:14] [SPEAKER_01]: But as we get clinical readouts, those readouts that we've done preclinically will have clinical

[00:32:18] [SPEAKER_01]: connectivity as well, and that will be very meaningful for us. We'll be able to sort of

[00:32:22] [SPEAKER_01]: further enhance the platform for predicting out and having validation that the readouts that we're

[00:32:26] [SPEAKER_01]: making, the predictions we're making in terms of thinking about preclinical safety correlate to

[00:32:32] [SPEAKER_01]: additionally and most importantly clinical safety. And that will be great in terms of extensions on

[00:32:37] [SPEAKER_01]: the platform. We're already I think about 100 times more efficient. We're going from Target

[00:32:41] [SPEAKER_01]: to Lead than others in the space. I think we're on track to add another order of magnitude to that

[00:32:47] [SPEAKER_01]: over the next couple of years. And if you think about what that means, really it puts us in a

[00:32:53] [SPEAKER_01]: place where our in silico approaches are actually starting to perform as well if not better than any

[00:33:00] [SPEAKER_01]: known preclinical assessment of a drug before it goes to the clinic. And just to give you context,

[00:33:05] [SPEAKER_01]: most drugs that go to the clinic, they start at the phase one trials which are mostly safety studies.

[00:33:11] [SPEAKER_01]: About 50% of those drugs fail in safety studies. So whatever we're doing to assess drugs from a

[00:33:16] [SPEAKER_01]: preclinical perspective, and I say we, just to be clear, I'm talking about the industry of Lark,

[00:33:21] [SPEAKER_01]: Crayon, and that's better than of course not being 50-50, but we have some room to do. But if our

[00:33:28] [SPEAKER_01]: actually allow us to do as well as that, think about the implications of what that means as you

[00:33:32] [SPEAKER_01]: think about translating genomic discoveries into new medicines. If you truly can engineer compounds

[00:33:38] [SPEAKER_01]: where you have the same probability of success going into the clinic as you would after spending

[00:33:44] [SPEAKER_01]: a decade doing preclinical research on sort of an activity than trial and error safety sort of

[00:33:51] [SPEAKER_01]: optimizations and then going into the clinic. We think that ultimately will have a huge potential

[00:33:56] [SPEAKER_01]: in terms of allowing us to really leverage the information that's come out of the post-genomic

[00:34:02] [SPEAKER_01]: world in a way that's quite exciting such that we can really sort of start treating the unmet

[00:34:08] [SPEAKER_01]: medical need that exists that we know how to correct. So that's, and I think where we are

[00:34:12] [SPEAKER_01]: with Crayon is we have a sort of a stepwise path to get there while also creating novel therapeutics

[00:34:19] [SPEAKER_01]: that meets unmet needs that can sort of fit into the development processes that exist today.

[00:34:24] [SPEAKER_01]: So we're systematically trying to walk through this process to really improve the efficiencies

[00:34:29] [SPEAKER_01]: of drug discovery and in the long run I think it'll be pretty exciting times.

[00:34:36] [SPEAKER_00]: Exciting times indeed. I can't thank you enough for taking the time to come on here and share your

[00:34:40] [SPEAKER_00]: insights with me but before I let you go, I'm going to ask you to leave one final gift to the

[00:34:45] [SPEAKER_00]: listeners and that is either a song that means something to you we can add to our Spotify

[00:34:50] [SPEAKER_00]: playlist, guilty pleasures are allowed or a book that we can add to an Amazon wish list. What would

[00:34:56] [SPEAKER_00]: you like to leave and why? Can I leave more than one book or do I have to leave more than one? Go

[00:35:00] [SPEAKER_01]: on, I'm feeling generous. Go on. What are you going to leave? Yes, the sort of I think about,

[00:35:06] [SPEAKER_01]: I try to read what I can but the two categories that I think are really interesting is that one

[00:35:10] [SPEAKER_01]: is contextualizing history. I found both and a lot of folks have seen this but Sapiens is being a

[00:35:15] [SPEAKER_01]: interesting book to contextualize where we are today but then another book which doesn't have

[00:35:20] [SPEAKER_01]: quite as much fanfare is the Ascent of Money. Both of those books are really interesting lenses to

[00:35:26] [SPEAKER_01]: look at history through which didn't necessarily reflect my academic work in history as I was going

[00:35:32] [SPEAKER_01]: through grade school and high school. Then from a company perspective, I found this book about Bell

[00:35:38] [SPEAKER_01]: Labs, The Idea Factory, quite fascinating about how the Bell Labs was able to cultivate a community

[00:35:45] [SPEAKER_01]: of really constant innovation and they set the stage for where we are today with modern electronics

[00:35:51] [SPEAKER_01]: and really the computer era and subsequently the internet age. Then I really enjoyed the books

[00:35:58] [SPEAKER_01]: around Amazon, both Everything Store and the sort of collection of shareable letters that Bezos put

[00:36:04] [SPEAKER_01]: together and invent and wonder. Then from a company creativity and culture perspective,

[00:36:11] [SPEAKER_01]: I thought Creativity Inc is pretty amazing book actually in the way that they do evaluations of

[00:36:18] [SPEAKER_01]: early movies to hone the story with everyone's feedback.

[00:36:23] [SPEAKER_00]: Awesome. Well, I will get all those books added to the Amazon wish list. For anyone listening,

[00:36:28] [SPEAKER_00]: just want to find out more information about you, your team or just dig a little bit deep on

[00:36:33] [SPEAKER_00]: anything we talked about today. Anywhere in particular you'd like to point them?

[00:36:37] [SPEAKER_01]: Now, LinkedIn is probably the best path you can find me if you search for Christopher Hart. You'll

[00:36:43] [SPEAKER_01]: find many Christopher Harts, but I'm the only Christopher Hart at Crayon. If you search for

[00:36:47] [SPEAKER_01]: Crayon and Christopher Hart, you'll probably find me. Then our website, crayonbio.com is

[00:36:52] [SPEAKER_01]: a good place to learn more about what we're doing.

[00:36:55] [SPEAKER_00]: Well, I've just loved chatting with you today about the role of AI, especially treating the

[00:37:01] [SPEAKER_00]: boy with the ultra rare disease there and also sharing insights on that intersection of AI,

[00:37:07] [SPEAKER_00]: medicine and drug development. They're both valuable as they are compelling, but just more

[00:37:12] [SPEAKER_00]: than anything, thank you so much for taking the time to sit down and share your story today.

[00:37:17] [SPEAKER_01]: Well, thank you, Neil. It's been great to chat.

[00:37:19] [SPEAKER_00]: So as we conclude our discussion with Chris, I think it's clear that this company is at the

[00:37:25] [SPEAKER_00]: a new era in medicine and by integrating AI into the very fabric of drug development,

[00:37:31] [SPEAKER_00]: they're not just making the process more efficient. They're changing how we approach

[00:37:35] [SPEAKER_00]: the treatment of rare and complex diseases. And the story of baby Leo's treatment is

[00:37:41] [SPEAKER_00]: for me an incredibly powerful example of what is possible when cutting edge technology

[00:37:47] [SPEAKER_00]: meets a deep commitment to patient safety and innovation.

[00:37:50] [SPEAKER_00]: And as I reflect on the conversation today, I think the potential for AI to transform not

[00:37:57] [SPEAKER_00]: just drug development timelines, but also the overall success and safety of new treatments

[00:38:02] [SPEAKER_00]: seems to be something incredibly magical happening here.

[00:38:06] [SPEAKER_00]: But how might these advancements impact the way that we approach healthcare,

[00:38:11] [SPEAKER_00]: the way we approach disease treatment in years to come?

[00:38:15] [SPEAKER_00]: These are a few of the things I'm going to be thinking about. Remember, email me

[00:38:19] [SPEAKER_00]: techblogwriteroutlook.com, Twitter, LinkedIn, Instagram, just at Neil C. Hughes.

[00:38:24] [SPEAKER_00]: Let me know what you're going to be thinking about after this conversation today.

[00:38:28] [SPEAKER_00]: I hope you leave this episode today with a renewed sense of possibilities that lie ahead

[00:38:33] [SPEAKER_00]: in the intersection of AI and medicine. We hear a lot of doom and gloom stories.

[00:38:38] [SPEAKER_00]: I'm hoping that we restored the balance in the universe a little today.

[00:38:42] [SPEAKER_00]: And on that positive note, I wish you well. Thank you for listening as always.

[00:38:47] [SPEAKER_00]: And hopefully you'll join me again tomorrow. You're all cordially invited.

[00:38:52] [SPEAKER_00]: So hopefully I'll speak with you all then. Bye for now.