Subscribe
Share
Share
Embed
Sounds of Science is a monthly podcast about beginnings: how a molecule becomes a drug, how a rodent elucidates a disease pathway, how a horseshoe crab morphs into an infection fighter. The podcast is produced by Eureka, the scientific blog of Charles River, a contract research organization for drug discovery and development. Tune in and begin the journey.
Brian Roche:
So being able to do it computerized, robotic, and then have readouts through the data system is certainly an advantage. You can filter through a whole host of compounds, a whole host of receptors, but there's still the receptors that you can't get to, and that's where, again, I think artificial intelligence is changing that for us even today
Mary Parker:
I'm Mary Parker and welcome to this episode of Sounds of Science. Today I am joined by Brian Roche, general manager of Charles River’s, Kansas site. We will explore how advanced technologies like automation, machine learning, and artificial intelligence are transforming drug development programs. Welcome Brian.
Brian Roche:
Thanks, Mary.
Mary Parker:
Good to have you here. So can you tell me a bit about yourself and your role at Charles River?
Brian Roche:
Sure. I joined Charles River in 2016 through the WIL Research acquisition. So I was in the Ashland, Ohio site, but prior to that I graduated from the Ohio State University with a degree in cardiovascular physiology, which coincided with regulations through the government based on safety pharmacology. So it thrust me into drug development early in my career, which I found was something that fit well. As the field grew, I grew, so I felt like I was a part of it from the start. One of the things about obviously safety pharmacology is it's the last phase of safety evaluations for drug development prior to going to human clinical trials. So how do you get involved early? How do you get more involved? Earlier in the process you apply the safety pharmacology tools earlier in drug development, which has kind of been my passion over my career. And then somewhere around 2022 I put in a strategy around disrupting drug development through utilizing automation, utilizing the advanced tools within artificial intelligence, big data to try to help drug development bring safer medicines to the patients that need them. our motto here at Charles River, and that led me to this site here in Kansas.
Mary Parker:
I mean, I guess that makes sense. The goal at the end of the day is to have drugs that are going to fail. You want them to fail as fast as possible so that they don't end up costing you more time and money if they're inevitably not going to be good. So yeah, applying those principles earlier seems like a good idea.
Brian Roche:
That's exactly right, and it's such a linear process. It's a process that's been in existence for a long time. We'll probably talk about the regulations that guide it. Those regulations are in place for the safety and efficacy of a product that moves through the drug development pipeline. But to your point, there is also oversensitivity around some of the assays that you might be tossing out safe compounds based on the indications that you see. So how do we really spend time, money, things in that space, as you mentioned, the fail quick because the business model is upside down. You need a product to get to the clinic to really fund your RD pipeline.
Mary Parker:
So what makes Kansas so unique?
Brian Roche:
Yeah, what's amazing, and that's the part, I mean I'll use the word disruption, but as a contract research organization, we offer a service and as that service has evolved over the decades, it's really become a partnership with clients. Clients have reduced their overhead internally. Maybe they're subject matter experts within a specific scientific discipline, and they rely on Charles River and the experts that we have. So we build our partnership that way. We now have the experts, we are leading the industry as far as how we assist drugs being developed. The uniqueness around the strategy in Kansas was just that how do you bridge discovery where you're testing a drug for efficacious and safety, but in that large gap where you want it to fail early, you want to be able to have a robust, quick, early dataset to make decisions. So how do you connect discovery to safety assessment?
How do you run it really quickly and nimbly? You start basically a site that's digitally enabled purpose-built just for this type of research. Now we'll probably get into the uphill climb on that because again, as I mentioned earlier, it's a linear process that's been in existence for a long time and you have to convince the regulators through data that the changes that you're making through innovation automation are reliable, robust, high quality, so that when they're reviewing it, they're comfortable looking at the changes. Things we talk about with alternative approaches to animals. How do you adopt such a thing as we have that science is out there today, but adoption through regulations is a lot harder because they have to be sure that the translation of what we're doing applies to the human nature of the evaluations in the clinic.
Mary Parker:
Before we get into the AI aspect of our conversation, I'd like to circle back to that regulatory aspect. Can you give any insight because I mean, basically end of the day, the FDA is an extremely conservative organization as it should be. It's all about patient safety. So how can you play into their sense of safety over something being, how do you play into their conservative mindset when you're trying to introduce a new type of technology or an alternative method?
Brian Roche:
I think you have to bring them into the conversation early and that willingness, that partnership around that will certainly help adopt it. The science drives the large wave, the wave of changeI think of 3D microfluidic organ systems. The science has been out there, we've picked it and pulled it apart early on. We're trying to make it more robust and each becomes more robust and eventually it will be something that gets in front of the regulators from a consort, so to speak, from a bunch of different data sets to say, see, here's where the changes are made. We're comfortable with what the changes are, and then how do we adopt such a thing Now, as you're right, it will take a long period of time. It's a slow process.
It's also then being able to demonstrate the data that comes out of some of the models, whether it be an artificial intelligence machine learning model, or whether it be actual data that you're collecting. Because what's hard about drug development from a linear process is the systems don't talk to each other. Nomenclature is different. The visual display of how you look at the data is different, but the regulators now know they have a send dataset, they have a standardized package that they look at from clinical aspect across the spectrum. That's really something we can build off of. How do we get similar data, visual tools that allow them to get to the individual data points to recreate and or investigate the data on their end and not just take a PDF table that's supplied to them as what they need to review.
Mary Parker:
So artificial intelligence, machine learning has been used a lot more and more in drug development. What advantages do artificial intelligence and machine learning bring to the drug development process?
Brian Roche:
Yeah, it's interesting because the artificial intelligence catchall term has under it machine learning, deep learning. So as long as we're talking about the right thing, and not just using the term artificial intelligence as a catchall, right, because they're specifically different, but understanding that artificial intelligence, the ability to utilize data algorithms to predict, to recreate, to in some cases create outcomes is really what we're talking about. So, on the forefront, if you break drug development down from in vitro single cell assessments, it's been going on for quite some time taking. If you think about being able to find a receptor from a therapy standpoint, being able to find something that matches to it, to then at least interact with it to do something about a disease state. In the drug development pipeline we have today, if you think about having a cell and then having the media to put on the cell, that process has to happen thousands and thousands of times, and it's a manual process.
So being able to do it computerized, robotic, and then have readouts through the data system is certainly an advantage. You can filter through a whole host of compounds, a whole host of receptors, but there's still the receptors that you can't get to, and that's where, again, I think artificial intelligence is changing that for us even today. So we'll start to see in the research, you'll see there'll be artificial intelligence similar to the music industry or writing a report for someone in college. There will be drugs that are developed based on computerized systems aiding its development throughout the process.
Brian Roche:
Because in vitro is just one piece of it, right? Then you have to, just like with anything else, you have to start looking at exposures and responses to the exposures as the drug moves into the in vivo setting. So that's where I think the horizon is for how we analyze data, but being able to utilize the data from discovery, and we'll probably talk a little bit about virtual control groups as an option of how you reduce the use of animals. The way you do that is you have to be able to retain the data from the efficacy studies that go into the predictive analytics or whatever analytics you're looking at. Then as you start to generate data, you can start to reduce things like, do you really need to run that study? Does it add any value to your decision-making properties today, humans might say, yes, I have to because I always do that and it gives me confidence that I can make this decision.
The computer might say, of the 10 studies you run, these three are really, they're providing you 99% of your answer. So that reduces studies, time animals if that happens to be an in vivo type study. But also, on that you can start to predict outcomes as they go out, meaning you start to see something based on a reference compound, thousands of reference compounds within a drug class to say, based on this response, this drug has this success rate and it has this pathway. These are the studies you should run to get there, or it's going to fail. It's going to fail in phase three, clinical development, we know this signature. By the time it gets through, we'll start to show that signal. You can't see it today. Again, human eyeballs on some data sets. The computer can get in there and look at those things and help us guide decisions as we move into the future. Again, all pathways that I think will change as drug development continues to evolve.
Mary Parker:
Those are some of the advantages. So what are some of the challenges associated with the use of AI?
Brian Roche:
The phrase that's used is bad data in, is bad data out. So certainly, it's about the data. It's also about how the drug development in the system that we utilize, that patented drugs that get to market that are licensed, not necessarily decrease the amount of sharing its intellectual property. There's certainly a consortium out there that work together. It's just how do you get the richness of the data? Can a single company generate enough data? One of the advantages of Charles River that we have is we have such a span of data. The virtual control groups, the number of vehicles run across species is hugely robust and it gives you that data confidence as you're starting to create models. But can companies do that? So, it's about the data. It's also about transparency of the data within the model, being able to really demonstrate to regulators or folks that are investing in what you're doing of how the algorithms work and what the data is that you get out of it.
I would think too, it's competitiveness. Not everyone can purchase things. We talk about automation, robotics, things that cost a lot of money, but over time will add such efficiency plays to the process that you have to be able to think long-term on where that investments go. But you also have to think about just a single molecule, a couple of folks trying to develop it, and large pharmaceutical companies that might have the capital to make those changes. So, it's partnering. It's partnering with companies that have already maybe established they're experts within the AI field. I think all of those need to be considered. And then lastly, it's about working with the regulators to be able to review that data and have confidence that we're moving in the right direction.
Mary Parker:
So, what about in terms of lead optimization, toxicology? Can you share your insights on how a data-driven approach that combines technology and automation can help that approach?
Brian Roche:
Yeah, and that's really the strategy here at the CRL Kansas site. We wanted to pull it out, redefine it, but it has to look like drug development today. That's how you convince sponsors, regulators, drug developers of why the change is needed, flexible, nimble, but robustness of the data. You also have to work on just being efficient in your workflow processes. So, an example of real life in the Kansas site is digital pathology. That laboratory is next generation equipment, next generation software. When I say laboratory information management systems that allow you to add the data integrity, but speed up the process, less time handling, less time checking because the computer has those checks built into it as it moves through the process, as well as just newer equipment that allows you to speed up the time it would take, for example, to scan an image at 40 to 80x resolution. Now that piece, because it's built off a digital pathology well accepted, is really robust.
Everyone's on board with, that's the direction where we need to go. But I think outside of that, even further, a lot has to happen before you get to digital pathology. So it's about managing the operations within an in vivo laboratory. It's about having automation within that to be able to schedule people, be able to schedule studies to be able to have equipment and resources available that can surge throughout a facility to accommodate the workload that's in it, as well as just processes that need to go through. If you think about the size of a vivarium being 10, 12 football fields and a study running on one end and samples being processed on the other, people have to walk back and forth daily to get samples to be able to process them. Some automation, robotics around moving those back and forth. Why then couldn't the robotics be centrifuging those samples as they're being moved back and forth? So, when they arrive at the facility, they can be processed out. We have to really focus on the science interpretation and the translation of the data.
Mary Parker:
Can you give some examples of how artificial intelligence and machine learning can be used for in vitro platforms?
Brian Roche:
I'm thinking about certainly rare and orphan disease status. Things that really affect as personalized medicine gets into it. We're really diving into those 8,000 diseases that really happen between birth and year two of life where the patient is at such a short period of time that you have to evaluate, and the number of patients is so small, it's not as if you have a lot of patients to pull from with regard to the treatment. But artificial intelligence, machine learning, the ability to identify those targets, be able to identify a molecule or a synthetic biologic that would target that, be able to do iterations of drug development that allow us to see what the changes might be, and it might be multiple treatments needed to know when the target itself is modulating so that the treatment is no longer as effective, but a secondary treatment on top of it would exemplify the finishing touches that might be needed.
I think those are key pieces as we move. I think for me it's about the development process. One of the examples I'll give is with machine learning and deep learning, you can train the algorithms to fill in a spot. So I'll talk about digital path again. You take a slice of a tissue, you put it on a slide under a microscope that we all did in biology in school, and you look at the changes. Digital scanning now allows that data to be digitized. So we can now utilize it in different ways than a hard microscopic platform with regard to glass. But there's also techniques where if it's out of focus during the scan, you have to re-scan it. Again, it's a reiteration in time, but with computer systems, you can actually predict the outcome of what the slide looks like from a small subset of the clear portion.
So again, an adoption of now not having to redo any work, the computer will just recreate the image. So we have to have confidence that that is truly the recreated image. If that's the case, you might not have to remove the tissue. You might just be able to biopsy it and utilize that small piece to then generate what the rest of it would look like additionally. Then you'd go, the next iteration would be then you just image the organ. You don't remove it at all or biopsy it at all. Then you're getting into the point again where you can assess it longitudinally and not just at the end of a certain period of treatment. All of these iterations, I think, help us get to the answers.
Mary Parker:
That also, of course, feeds into the idea that the data that the machine is fed has to be perfect. It has to be really good if the machine is going to be creating or even recreating anything. It has to basing its judgment on data that is clean and polished and accurate.
Brian Roche:
Exactly right, and well characterized. Yep.
Mary Parker:
How are in vitro and in silico innovations being used to advance alternative methods?
Brian Roche:
There's so much science out there that's been going on for the last couple of decades that it's really coming to a head. You'll see across a lot of the scientific meetings that we attend on the international scale, really focusing on the development of these systems as well as reducing the use of animals through the alternative approaches. It really, to me hinges on understanding what's acceptable. Again, through the drug development review process. I think the more we learn about programs that have been approved using alternative methods will continue to drive it. One of those pivot points we talked about that makes folks a little bit nervous is if you alter from the standard and you get to the review process and you have to go backwards, you're slowing down your development, you're increasing costs, but you're also reducing the time it would take to get that drug to a patient.
So understanding what is acceptable and the programs that are utilizing it is important. Certainly, if we start to utilize more of the in silico programming outcomes, predictive models, that those things are transparent within the drug development portfolio that's submitted. So, we know, again, what was the pathway that we followed to get that drug approved to get to the clinic. So, if I get back to your question, it's really about utilizing the algorithms. One of the examples that I can give you from safety pharmacology is how do you go from a single cell utilizing an action potential within your heart to arrhythmia in the human? And you do that by getting the outputs from the single cell, utilizing your in silico algorithms to recreate it from potentially what we would utilize in that in vitro system to a cardiomyocyte to then a fully formed heart is part of where the advances in silico can be adapted from.
Mary Parker:
So what other progress has Charles River made in intersecting our scientific and technological innovations?
Brian Roche:
So that's where I think we're leading. Charles River has been so far ahead with their partnerships, the way they work, through the whole host of services available from discovery through the clinic. We're connected through bioanalytical data, right? The exposures that we see in the clinic and what we're seeing pre-clinically. How do we connect things that we see pre-clinically to translation to the clinic. Again, all important factors. We're not just doing a task and moving it to the next task. We're developing drugs and partnering with drug developers to do that. So again, transparency across the spectrum on what they're seeing in the clinic and what we're seeing pre-clinically is key. I think utilizing the partnerships that we have for machine learning, artificial intelligence, those are big points We've talked about. Again, the data sets that they utilize, we have to, as a company, the platforms that we collect that data on, align the virtual control groups is allowing us to do that.
As you mentioned earlier, a lot of the in-life data comes from a tool called Provantis, but then a lot of the pharmacodynamic data that's continuous heart rate, blood pressure, things that you measure in the clinic aren't loaded through that software. So you have to be able to be able to pull that out from the different sources to be able to utilize that within those models. But you also have to convince regulators as we talked about. So you want to find something that does translate as the tool so that you can build off of it. It's not just about that one tool. It's about the holistic approach. To me as a researcher, as I look at data, you think about clinical pathology, everyone goes to the doctor, you get blood drawn and you get a whole host of outputs that says, yes, all those indicators are within normal limits.
But if you're looking across dose groups, across multiple studies, the human looking at all that data, you group it into means, and then you look at the mean change based on dose response, and then you make decisions about the biological relevance and your decisions are within historical controls. But it's, again, it's a judgment around what you're used to seeing your scientific integrity around what you're used to as far as experimental designs to say that this is meaningful and this is not utilizing tools that can look at all the data points, outliers are not, and to see is the outlier meaningful? We would say if it's one of 10, it's probably an airer signal, but is it the one that's telling us more about what's going on? Can the computer systems then go in there and look at that outlier and look across the spectrum to see not just that outlier, but the other data points for that subject to see what's that signal? Does that animal or human or cell have the polymorphism? That does lead us to understand that that's going to create an issue when we start giving that to thousands of patients when you start writing scripts. I think those are all important things to consider.
Mary Parker:
Yeah, after all, one data point out of 10 could be nothing, could be a mistake, could be a random fluke, or it could represent 10% of your patient pool. So yeah, that's an important thing to distinguish,
Brian Roche:
And computers can allow us to do that millions of data points.
Mary Parker:
Yeah. So speaking of taking a holistic approach and partnerships, I assume that when you're working with machine learning partnerships and collaborations with clients and with other companies is even more important because the more data the better.
Brian Roche:
No, it's a really good point. And building off of the previous question. So, one real life example is even I'm going to stick with digital pathology. The pathologist looks at the scan, looks at the staining on the scan to say that this change has happened. They quantify the change, and then they write that up in concert with other things that happen on study. But there are tools that can be used to highlight a lot of the times you're looking through slides or scans that are negative. So, there are tools, artificial intelligence tools today that apply that tell you, look at this slide. These are the four slides that had changes, and they put a mark on them to say, here's where you should be looking. Now, it doesn't remove the pathologist, the expert from looking at it, but it helps us narrow down where to look.
Additionally, that next phase of development, and you're talking about it here, is how do you partner with folks to move to the next generation? Because there are companies that sell a service, but then there are some that will partner with you to work through the next generation of the service. How does it fit your need? And that's key in selecting your partners because we don't know what the answers are for tomorrow, but we know this technology match with this technology will get us there, but we have to work together to get there. So, the example I'll give again is staining of a slide. It happens manually. Chemicals are used, resources are used, and of course you're exposing humans to the chemicals. So, there are virtual scanning tools available, partners available that have virtual AI staining. So again, working with those companies to qualify those systems, increase their capacity, and then utilize that within studies to create consortia response through the regulators is another way to remove a piece that happens today that's manual and allow us to digitize and not just utilize more tissues to make different scans, but utilize the original tissue to scan stains of whatever interest you're looking for within the tissue itself.
It's all where the science is going.
Mary Parker:
Well, as we kick off 2025, what excites you the most about the year ahead? And you can say digital pathology.
Brian Roche:
Again, where we are within lead optimization toxicology through the site in Kansas, we've now been open for over a year. We've had many opportunities to speak about what we're doing here, which means we've had more opportunity to talk to folks about the change in drug development that we're trying to institute. The science is robust. The difference is knowing that it's here and it's networked globally throughout Charles River, to tap into any expertise that's needed as the drug moves through the platform is key. So I feel like moving into 2025, we're poised with a lot more recognition of what we're trying to do. We're trying to meet the needs that drug developers need today, which is speed, flexibility, nimbleness, robust data, high quality data to make decisions in this early space.
Mary Parker:
Well, and also as you mentioned before, with digital pathology, being able to use one sample in many different ways, doing more with less in the current economy is a valuable thing.
Brian Roche:
Exactly. Right. And we know utilizing the data sets from the last five, 10 years of studies, how to optimize the study as we talked about, reduce the use of animals to get the answer that you need. We've truncated the process so that we're working more efficiently through it, but also being robust that you can get to the send dataset that you might need when you submit that compound within your package.
Mary Parker:
And finally, in your opinion, what does the future hold for the relationship between advanced technologies and drug development? Is there a ceiling that we might hit at some point?
Brian Roche:
Yeah, that's interesting. I'm sure we're going to hit multiple ceilings, and I'm hoping they're not closed roofs if that makes sense. I hope that it allows us to get to the next iteration because that's what helps us get to the truth, right? It's proving what you've learned, even though it might be a negative to move in another direction to get there. I think the novel approaches the science behind that. The adoption is really starting to catch on. Being able to translate those things across the spectrum, across the scales is key. As we continue to develop that and getting to a point where the regulators and drug developers are together on how the change is going to be made. Now, as I've stated, it's going to take time and it's going to take data to convince, but it's also working with, what are we talking about? We're talking about patients that need these drugs, that need these treatments. I feel like we're at a point that we've never been at before with focus on that target.
Mary Parker:
Totally agree. Thank you so much, Brian, for joining us for this discussion. It's been really interesting.
Brian Roche:
Yeah, thank you. I appreciate the opportunity.
Mary Parker:
Thank you to Brian Roche, General Manager of Charles River’s Kansas site. Stay tuned for the next episode of Sounds of Science. Until then, you can subscribe to Sounds of Science on Apple Podcasts, Spotify, Stitcher, or wherever you get your podcasts. Thanks for listening.